Quantitative Fund Management

Quantitative Fund Management

CHAPMAN & HALL/CRC Financial Mathematics Series Aims and scope: The field of financial mathematics forms an ever-expanding slice of the financial sector. This series aims to capture new developments and summarize what is known over the whole spectrum of this field. It will include a broad range of textbooks, reference works and handbooks that are meant to appeal to both academics and practitioners. The inclusion of numerical code and concrete real-world examples is highly encouraged.

Series Editors M.A.H. Dempster Centre for Financial Research Judge Business School University of Cambridge

Dilip B. Madan Robert H. Smith School of Business University of Maryland

Rama Cont Center for Financial Engineering Columbia University New York

Published Titles American-Style Derivatives; Valuation and Computation, Jerome Detemple Credit Risk: Models, Derivatives, and Management, Niklas Wagner Engineering BGM, Alan Brace Financial Modelling with Jump Processes, Rama Cont and Peter Tankov An Introduction to Credit Risk Modeling, Christian Bluhm, Ludger Overbeck, and Christoph Wagner Introduction to Stochastic Calculus Applied to Finance, Second Edition, Damien Lamberton and Bernard Lapeyre Numerical Methods for Finance, John A. D. Appleby, David C. Edelman, and John J. H. Miller Portfolio Optimization and Performance Analysis, Jean-Luc Prigent Quantitative Fund Management, M. A. H. Dempster, Georg Pflug, and Gautam Mitra Robust Libor Modelling and Pricing of Derivative Products, John Schoenmakers Structured Credit Portfolio Analysis, Baskets & CDOs, Christian Bluhm and Ludger Overbeck Understanding Risk: The Theory and Practice of Financial Risk Management, David Murphy

Proposals for the series should be submitted to one of the series editors above or directly to: CRC Press, Taylor & Francis Group 4th, Floor, Albert House 1-4 Singer Street London EC2A 4BQ UK

Quantitative Fund Management

Edited by

M. A. H. Dempster, Georg Pflug, and Gautam Mitra

Chapman & Hall/CRC Taylor & Francis Group 6000 Broken Sound Parkway NW, Suite 300 Boca Raton, FL 33487-2742 © 2009 by Taylor & Francis Group, LLC Chapman & Hall/CRC is an imprint of Taylor & Francis Group, an Informa business No claim to original U.S. Government works Printed in the United States of America on acid-free paper 10 9 8 7 6 5 4 3 2 1 International Standard Book Number-13: 978-1-4200-8191-6 (Hardcover) This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may rectify in any future reprint. Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, please access www.copyright.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users. For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation without intent to infringe. Library of Congress Cataloging-in-Publication Data Dempster, M. A. H. (Michael Alan Howarth), 1938Quantitative fund management / Michael Dempster, Georg Pflug, and Gautam Mitra. p. cm. Includes bibliographical references and index. ISBN 978-1-4200-8191-6 (alk. paper) 1. Portfolio management--Mathematical models. 2. Investment analysis--Mathematical models. I. Pflug, Georg Ch., 1951- II. Mitra, Gautam, 1947- III. Title. HG4529.5.D465 2008 332.63’2042--dc22 Visit the Taylor & Francis Web site at http://www.taylorandfrancis.com and the CRC Press Web site at http://www.crcpress.com

2008014075

Contents Editors

vii

Contributors

ix

Introduction

xiii

PART 1 j Dynamic Financial Planning CHAPTER 1 j Trends in Quantitative Equity Management: Survey Results CHAPTER 2 j Portfolio Optimization under the Value-at-Risk Constraint CHAPTER 3 j Dynamic Consumption and Asset Allocation with Derivative Securities CHAPTER 4 j Volatility-Induced Financial Growth CHAPTER 5 j Constant Rebalanced Portfolios and Side-Information CHAPTER 6 j Improving Performance for Long-Term Investors: Wide Diversification, Leverage and Overlay Strategies CHAPTER 7 j Stochastic Programming for Funding Mortgage Pools CHAPTER 8 j Scenario-Generation Methods for an Optimal Public Debt Strategy CHAPTER 9 j Solving ALM Problems via Sequential Stochastic Programming CHAPTER 10 j Designing Minimum Guaranteed Return Funds PART 2 j Portfolio Construction and Risk Management CHAPTER 11 j DC Pension Fund Benchmarking with Fixed-Mix Portfolio Optimization CHAPTER 12 j Coherent Measures of Risk in Everyday Market Practice CHAPTER 13 j Higher Moment Coherent Risk Measures CHAPTER 14 j On the Feasibility of Portfolio Optimization under Expected Shortfall CHAPTER 15 j Stability Analysis of Portfolio Management with Conditional Value-at-Risk

1 3 17 43 67 85 107 129 175 197 223 245 247 259 271 299 315 v

vi j CONTENTS

CHAPTER 16 j Stress Testing for VaR and CVaR CHAPTER 17 j Stable Distributions in the Black
Litterman Approach to Asset Allocation CHAPTER 18 j Ambiguity in Portfolio Selection CHAPTER 19 j Mean-Risk Models Using Two Risk Measures: A Multi-Objective Approach CHAPTER 20 j Implied Non-Recombining Trees and Calibration for the Volatility Smile Index

337

/

359 377 393 425 451

Editors Dr. M.A.H. Dempster was educated at Toronto (BA), Oxford (MA), and Carnegie Mellon (MS, PhD) Universities and is currently professor emeritus at the Statistical Laboratory, Centre for Mathematical Sciences, in the University of Cambridge. Previously he was director of the Centre for Financial Research, director of research and director of the doctoral programme at the Judge Business School. He was also professor of mathematics and director of the Institute for Studies in Finance in the University of Essex, R. A. Jodrey Research Professor of Management and Information Sciences and professor of Mathematics, Statistics and Computing Science at Dalhousie University and fellow and University Lecturer in Industrial Mathematics at Balliol College, Oxford. Since 1996 he has been the managing director of Cambridge Systems Associates Limited, a financial services consultancy and software company. His teaching experience is in a broad range of mathematical, decision, information and managerial sciences, and he has supervised over 40 doctoral students in these fields. He has held visiting positions at Stanford, Berkeley, Princeton, Toronto, Rome and IIASA. He has also been a consultant to numerous leading international commercial and industrial organizations and several governments and is in demand as a speaker and executive educator around the world. As well as mathematical and computational finance and economics, his present research interests include optimization and nonlinear analysis, stochastic systems, algorithm analysis, decision support and applications software. He is author of over 100 published research articles and reports and is author, editor or translator of 11 books including Introduction to Optimization Methods (with P. R. Adby), Large Scale Linear Programming (2 vols., with G. B. Danzig and M. Kallio), Stochastic Programming, Deterministic and Stochastic Scheduling (with J. K. Lenstra and A. H. G. Rinnooy Kan), Mathematics of Derivative Securities (with S. R. Pliska) and Risk Management: Value at Risk and Beyond. He is founding joint Editor-in-Chief of Quantitative Finance with J. Doyne Farmer and presently shares this position with J.-P. Bouchaud. He was formerly on the editorial boards of the Review of Economic Studies, Journal of Economic Dynamics and Control, Mathematical Finance and Computational Economics and is currently an associate editor of Stochastics, Computational Finance and the Journal of Risk Management in Financial Institutions. He received the D. E. Shaw Best Paper Award at Computational Intelligence in Financial Engineering 1999 and became an honorary fellow of the UK Institute of Actuaries in 2000. In 2004 The Mathematical Programming Society recognized him as a Pioneer of Stochastic Programming. vii

viii j EDITOR

Dr. Gautam Mitra qualified as an electrical engineer at Jadavpur University, Kolkata. He then joined London University, where he completed first an MSc followed by a PhD in computer methods in operational research. Dr. Mitra joined the faculty of Mathematics, Statistics and OR at Brunel University in 1974; since 1988 he has held a chair of Computational Optimisation and Modelling, and between 1990
2001 he was head of the Department of Mathematical Sciences. In 2001, under the strategic research initiative of Brunel University, Professor Mitra established Centre for the Analysis of Risk and Optimisation Modelling (CARISMA) and has led it as its Director. Professor Mitra’s research interests cover mathematical optimisation and, more recently, financial modelling and risk quantifications. In the year 2003 Brunel University, in recognition of his academic contributions, honoured him as a distinguished professor. Professor Mitra, prior to his academic career at Brunel University, worked with SCICON, SIA and ICL; in all these organisations he was involved in the (software) development of mathematical programming optimisation and modelling systems. He was the director of advanced NATO Advanced Study Research Institute, which took place in Val-d-Isere (1987) on the topic of Mathematical Models for Decision Support. Professor Mitra has also had substantial involvement with industry-based projects in his role as a director of UNICOM Seminars and OptiRisk Systems. He is the author of two textbooks, four edited books and over ninety-five journal papers. His Web sites are http:// www.carisma.brunel.ac.uk and http://www.optirisk-systems.com /

Dr. Georg Pflug studied law, mathematics and statistics at the University of Vienna. He was assistant professor at the University of Vienna, professor at the University of Giessen, Germany and is currently full professor at the University of Vienna and Head of the Computational Risk Management Group. He will be Dean of the Faculty of Business, Economics and Statistics in the period 2008
2010. Georg Pflug held visting positions University of Bayreuth, Michigan State University, University of California at Davis, Universite´ de Rennes, Technion Haifa and Princeton University. He is also part time research scholar at the International Institute of Applied Systems Analysis, Laxenburg, Austria. He is currently editor-in-chief of Statistics and Decisions and associate editor of Stochastic Programming Electronic Publication Series, Central European Journal of Operations Research, Austrian Journal of Statistics, Mathematical Methods of OR, Computational Optimization and Applications, and Computational Management Science. Georg Pflug is author of 4 books, editor of 6 books, and has written more than 70 publications in refereed journals, such as: Annals of Statistics, Annals of OR, Probability Theory, J. Statist. Planning and Inference, J. ACM, Parallel Computing, The Computer Journal, Math. Programming, Mathematics of Optimization, SIAM J. on Optimization, Computational Optimization and Applications, J. Applied Probability, Stoch. Processes and Applications, Graphs and Combinatorics, J. Theoretical Computer Science, Journal of Banking and Finance, Quantitative Finance and others. /

Contributors Carlo Acerbi Abaxbank Milano, Italy Massimo Bernaschi Istituto per le Applicazioni del Calcolo ‘Mauro Picone’—CNR Roma, Italy Marida Bertocchi Department of Mathematics, Statistics, Computer Science and Applications Bergamo University Bergamo, Italy Maya Briani Istituto per le Applicazioni del Calcolo ‘Mauro Picone’—CNR Roma, Italy Chris Charalambous Department of Business Administration University of Cyprus Nicosia, Cyprus

Eleni D. Constantinide Department of Business Administration University of Cyprus Nicosia, Cyprus Kenneth Darby-Dowman The Centre for the Analysis of Risk and Optimisation Modelling (CARISMA) Brunel University London, U.K. M. A. H. Dempster Centre for Financial Research, Judge Business School University of Cambridge & Cambridge Systems Associates Ltd. Cambridge, U.K. Gabriel Dondi ETH Zurich Zurich, Switzerland and swissQuant Group AG Zurich, Switzerland

Nicos Christofides Centre for Quantitative Finance Imperial College London, U.K.

Jitka Dupacˇova´ Faculty of Mathematics and Physics, Department of Probability and Mathematical Statistics Charles University Prague, Czech Republic

Stefano Ciliberti CNRS Universite´ Paris Sud Orsay, France

Igor V. Evstigneev School of Economic Studies University of Manchester Manchester, U.K. ix

x j CONTRIBUTORS

Frank J. Fabozzi Yale School of Management Yale University New Haven, Connecticut E. Fagiuoli Dipartimento di Informatica Sistemistica e Comunicazione, Universita` degli Studi di Milano-Bicocca Milano, Italy Sergio Focardi Yale School of Management Yale University New Haven, Connecticut Hans P. Geering ETH Zurich Zurich, Switzerland M. Germano Pioneer Investment Management Ltd. Dublin, Ireland Rosella Giacometti Department of Mathematics, Statistics, Computer Science and Applications Bergamo University Bergamo, Italy Florian Herzog ETH Zurich Zurich, Switzerland and swissQuant Group AG Zurich, Switzerland

Yuan-Hung Hsuku Department of Financial Operations National Kaohsiung First University of Science and Technology Taiwan, Republic of China Gerd Infanger Department of Management Science and Engineering Stanford University Stanford, California and Infanger Investment Technology, LLC Mountain View, California Caroline Jonas The Intertek Group Paris, France Michal Kaut Molde University College Molde, Norway Simon Keel ETH Zurich Zurich, Switzerland and swissQuant Group AG Zurich, Switzerland Imre Kondor Collegium Budapest Budapest, Hungary Pavlo A. Krokhmal Department of Mechanical and Industrial Engineering The University of Iowa Iowa City, Iowa

CONTRIBUTORS

j

Spiros H. Martzoukos Department of Business Administration University of Cyprus Nicosia, Cyprus

Traian A. Pirvu Department of Mathematics The University of British Columbia Vancouver, British Columbia, Canada

E. A. Medova Centre for Financial Research, Judge Business School University of Cambridge & Cambridge Systems Associates Ltd. Cambridge, U.K.

Jan Polı´vka Faculty of Mathematics and Physics, Department of Probability and Mathematical Statistics Charles University Prague, Czech Republic

Marc Me´ zard CNRS Universite´ Paris Sud Orsay, France Gautam Mitra The Centre for the Analysis of Risk and Optimisation Modelling (CARISMA) Brunel University London, U.K. John M. Mulvey Bendheim Center of Finance Princeton University Princeton, New Jersey Marco Papi Istituto per le Applicazioni del Calcolo ‘Mauro Picone’—CNR Roma, Italy Georg Pflug Department of Statistics and Decision Support Systems University of Vienna Vienna, Austria

xi

Svetlozar T. Rachev School of Economics and Business Engineering University of Karlsruhe Karlsruhe, Germany and Department of Statistics and Applied Probability University of California Santa Barbara, California M. I. Rietbergen Securitisation & Asset Monetisation Group Morgan Stanley London, U.K. Diana Roman The Centre for the Analysis of Risk and Optimisation Modelling (CARISMA) Brunel University London, U.K. F. Sandrini Pioneer Investment Management Ltd. Dublin, Ireland

xii j CONTRIBUTORS

Klaus R. Schenk-Hoppe´ Leeds University Business School and School of Mathematics University of Leeds Leeds, U.K. Lorenz M. Schumann swissQuant Group AG Zurich, Switzerland M. Scrowston Pioneer Investment Management Ltd. Dublin, Ireland F. Stella Dipartimento di Informatica Sistemistica e Comunicazione, Universita` degli Studi di Milano-Bicocca Milano, Italy Cenk Ural Lehman Brothers New York, New York and Blackrock, Inc. New York, New York A. Ventura Dipartimento di Informatica Sistemistica e Comunicazione, Universita` degli Studi di Milano-Bicocca Milano, Italy Davide Vergni Istituto per le Applicazioni del Calcolo ‘Mauro Picone’—CNR Roma, Italy

Hercules Vladimirou Centre for Banking and Financial Research, School of Economics and Management University of Cyprus Nicosia, Cyprus Stein W. Wallace Molde University College Molde, Norway David Wozabal Department of Statistics and Decision Support Systems University of Vienna Vienna, Austria Stavros A. Zenios Centre for Banking and Financial Research, School of Economics and Management University of Cyprus Nicosia, Cyprus N. Zhang ABN Amro London, U.K. Zhuojuan Zhang Blackrock, Inc. New York, New York

Introduction to Quantitative Fund Management

A

on stochastic programming held at the University of Arizona in October 2004, it was observed that the fund management industry as a whole was far from the leading edge of research in financial planning for asset allocation, asset liability management, debt management and other financial management problems at the strategic (long term) level. This gap is documented in the timely survey of quantitative equity management by Fabozzi, Focardi and Jonas which forms the first chapter of this book. It was therefore agreed to bring out a special issue of Quantitative Finance to partially address the imbalance between research and practice by showcasing leading edge applicable theory and methods and their use for practical problems in the industry. A call for papers went out in August and October of 2005. As an outcome of this, we were able to compile a first special issue with the papers forming the ten chapters in Part 1 of this book. In fact, the response to the call was so good that a second special issue focusing on tactical financial planning and risk management is contained in the ten chapters of Part 2. Taken together, the twenty chapters of this volume constitute the first collection to cover quantitative fund management at both the dynamic strategic and one period tactical levels. They consider optimal portfolio choice for wealth maximization together with integrated risk management using axiomatically defined risk measures. Solution techniques considered include novel applications to quantitative fund management of stochastic control, dynamic stochastic programming and related optimization techniques. A number of chapters discuss actual implemented solutions to fund management problems including equity trading, pension funds, mortgage funding and guaranteed investment products. All the contributors are well known academics or practitioners. The remainder of this introduction gives an overview of their contributions. In Part I of the book on dynamic financial planning the survey by Fabozzi et al. (Chapter 1) finds that, at least in the equity world, the interest in quantitative techniques is shifting from basic Markowitz mean-variance portfolio optimization to risk management and trading applications. This trend is represented here with the chapter by Fagiuoli, Stella and Vetura (Chapter 5). The remaining chapters in Part 1 cover novel aspects of lifetime individual consumption investment problems, fixed mix portfolio rebalancing allocation strategies (including Cover-type universal portfolios), debt management for T THE TENTH TRIENNIAL INTERNATIONAL CONFERENCE

xiii

xiv j INTRODUCTION TO QUANTITATIVE FUND MANAGEMENT

funding mortgages and national debt, and guaranteed return fund construction. Of the ten chapters in Part 1, one is the mentioned survey, three are theoretical, two concern proofs of concept for practical trading or fund management strategies and the remaining four concern real-world implementations for major financial institutions. Chapter 2 by Pirvu expands on the classical consumption investment problem of Merton to include a value-at-risk constraint. The portfolio selection problem over a finite horizon is a stochastic control problem which is reduced to pathwise nonlinear optimization through the use of the stochastic Pontryagin maximum principal. Numerical results are given and closed form solutions obtained for special cases such as logarithmic utility. The third chapter by Hsuku extends the classical Merton problem in a different direction to study the positive effects of adding derivatives to investors’ choices. The model utilizes a recursive utility function for consumption and allows predictable variation of equity return volatility. Both of these theoretical studies concern realistically incomplete markets in which not all uncertainties are priced. The next three chapters mainly treat variants of the fixed-mix rebalance dynamic asset allocation strategy. The first of these (Chapter 4) by Dempster, Evstigneev and SchenkHoppe´ shows under very general stationary ergodic return assumptions that such a strategy, which periodically rebalances a portfolio to fixed proportions of the current portfolio value, grows exponentially on almost every path even in the presence of suitable transactions costs. Chapter 5 in this group by Fagiuoli, Stella and Ventura develops, and tests on stock data from four major North American indices, an online algorithm for equity trading based on Cover’s non-parametric universal portfolios in the situation when some market state information is also available. Chapter 6 by Mulvey, Ural and Zhang discusses return enhancing additions to both fixed mix rebalance strategies and optimal dynamic allocation strategies obtained by dynamic stochastic programming in the context of work for the U.S. Department of Labor. In particular, positive return performance is demonstrated from diversification to non-traditional asset classes, leverage, and overlay strategies which require no investment capital outlay. The next two chapters concern debt management problems which use dynamic stochastic programming to optimally fund mortgage lending and government spending requirements respectively. These are asset liability management problems in which assets are specified and decisions focus on liabilities, namely, when and how to issue bonds. The first, Chapter 7 by Infanger, is an exemplary study conducted for Freddie Mac which shows that significant extra profits can be made by employing dynamic models relative to static Markovitz mean-variance optimization or traditional duration and convexity matching of assets (mortgage loans) and liabilities (bonds). In addition, efficient out-ofsample simulation evaluation of the robustness of the recommended optimal funding strategies is described, but not historical backtesting. Chapter 8 by Bernaschi, Briani, Papi and Vergni concentrates on yield curve modelling for a dynamic model for funding Italian public debt by government bond issuance. The idea of this contribution, important in an EU context, is to model the basic ECB yield curve evolution together with an orthogonal national idiosyncratic component.

INTRODUCTION TO QUANTITATIVE FUND MANAGEMENT

j

xv

The last two chapters of Part 1 describe the use of dynamic stochastic programming techniques to design guaranteed return pension funds which employ dynamic asset allocations to balance fund return versus guarantee shortfall. Chapter 9 by Hertzog, Dondi, Keel, Schumann and Geering treats this asset liability management problem using a deterministic evolution of the guarantee liability, while Chapter 10 by Dempster, Germano, Medova, Rietbergen, Sandrini and Scrowston treats guarantee shortfall with respect to a stochastic liability which is evaluated from the forward ECB yield curve simulation used to price bonds in the dynamic portfolio. Both chapters employ historical backtesting of their models for respectively a hypothetical Swiss pension fund and (a simplified version of) actual funds backing guaranteed return products of Pioneer Investments. Taken together, the ten chapters of Part 1 give a current snapshot of state-of-the-art applications of dynamic stochastic optimization techniques to long term financial planning. These techniques range from new pathwise (Pirvu) and standard dynamic programming (Hsuku) methods of stochastic control, through sub-optimal, but easily understood and implemented policies (Dempster et al., Fagiouli et al., Mulvey et al.) to dynamic stochastic programming techniques involving the forward simulation of many risk factors (Mulvey et al., Infanger, Bernaschi et al., Hertzog et al., Dempster et al.). Although there is currently widespread interest in these approaches in the fund management industry, more than a decade after their commercial introduction they are still in the early stages of adoption by practitioners, as the survey of Fabozzi et al. shows. This volume will hopefully contribute to the recognition and wider acceptance of stochastic optimization techniques in financial practice. Part 2 of this volume on portfolio construction and risk management concerns the tactical level of financial planning. Most funds, with or without associated liabilities—and explicitly or implicitly—employ a three level hierarchy for financial planning. The top strategic level considers asset classes and risk management over longer term horizons and necessarily involves dynamics (the topic of Part 1). The middle tactical level of the financial planning hierarchy concerns portfolio construction and risk management at the individual security or fund manager level over the period up to the next portfolio rebalance. This is the focus of the ten contributions of the second part of the book. The third and bottom operational level of the financial planning hierarchy is actual trading which, with the rise of hedge funds, and as the survey of quantitative equity management by Fabozzi et al. in Chapter 1 demonstrates, is becoming increasingly informed by tactical models and considerations beyond standard Markowitz mean-variance optimization (MVO). This interaction is the evident motivation for many of the chapters in Part 2 with their emphasis on non-Gaussian returns, new risk-return tradeoffs and robustness of benchmarks and portfolio decisions. The first two chapters are based on insights gained from actual commercial applications, while of the remaining eight chapters all but one, which is theoretically addressing an important practical issue, test new theoretical contributions on market data. Another theme of all the contributions in this part is that their concern is with techniques which are scenario—rather than analytically—based (although the purely theoretical chapter uses a limiting analytical approximation). This

xvi j INTRODUCTION TO QUANTITATIVE FUND MANAGEMENT

theme reflects the necessity for nontrivial computational approaches when the classical independent Gaussian return paradigm is set aside in favour of non-equity instruments and shorter term (e.g. daily or weekly) returns. The first chapter of Part 2, Chapter 11 by Dempster, Germano, Medova, Rietbergen, Sandrini, Scrowston and Zhang treats the problem of benchmarking fund performance using optimal fixed mix rebalancing strategies (a theme of Part 1) and tests it relative to earlier work on optimal portfolios for guaranteed return funds described in Chapter 10. Chapter 12 by Acerbi provides a timely and masterful survey of the recent literature on coherent risk measures, including practical linear programming models for portfolios constructed by their minimization. This theme is elaborated further by Krokhmal in Chapter 13 which treats higher moment coherent risk measures. It examines their theoretical properties and performance when used in portfolio construction relative to standard mean variance and expected shortfall conditional value at risk (CVaR) optimization. The next three chapters treat the robustness properties of the numerical minimization of CVaR using linear programming as employed in practice, for example, for bond portfolios. The first, Chapter 14 by Ciliberti, Kondor and Mezard, uses limiting continuous approximations suggested by statistical physics to define a critical threshold for the ratio of the number of assets to the number of historical observations beyond which the expected shortfall (CVaR) risk measure is not well-defined—a phase-change phenomenon first noted by Kondor and co-authors. Next Kaut, Vladimirou, Wallace and Zenios examine in Chapter 15 the stability of portfolio solutions to this problem with respect to estimation (from historical data) errors. They conclude that sensitivity to estimation errors in the mean, volatility, skew and correlation all have about the same non-negligible impact, while error in kurtosis has about half that of the other statistics. Finally, Chapter 16 by Dupacˇova and Polı´vka discusses stress-testing the CVaR optimization problem using the contamination scenario technique of perturbation analysis. They also show that similar techniques may be applied to the minimal analytical value at risk (VaR) problem for the Gaussian case, but are not applicable to the corresponding historical scenario based problem. The next group of three chapters extend the treatment of portfolio construction and risk management beyond the usual simple tradeoff of volatility risk and return embodied in MVO. Chapter 17 by Giacometti, Bertocchi, Rachev and Fabozzi shows that the BlackLitterman Bayesian approach to portfolio construction, incorporating both market and practitioner views, can be extended to Student-t and stable return distributions and VaR and CVaR risk measures. Pflug and Wozabal consider in Chapter 18 the robust optimization problem of finding optimal portfolios in the Knightian situation when the distributions underlying returns are not perfectly known. They develop and test an algorithm for this situation based on two level convex optimization. In the last chapter in this group, Chapter 19, Roman, Darby-Dowman and Mitra consider the multi-objective problem of simultaneously trading off expected return with two risk measures based on variance and expected shortfall (CVaR). In tests with FTSE 100 index securities they find that an optimal balance with the two risk measures dominates those using either alone.

INTRODUCTION TO QUANTITATIVE FUND MANAGEMENT

j

xvii

The final chapter in Part 2, Chapter 20 by Charalambous, Christofides, Constantinide and Martzoukos, treats the basic requirement for pricing exotic and over-the-counter options—fitting vanilla option market price data—using non-recombining (binary) trees, a special case of the multi-period scenario trees used in Part 1 for strategic portfolio management. The authors’ approach dominates the usual recombining tree (lattice) in that it can easily handle transactions costs, liquidity constraints, taxation, non-Markovian dynamics, etc. The authors demonstrate its practicality using a penalty method and quasiNewton unconstrained optimization and its excellent fit to the volatility surface—crucial for hedging and risk control. The ten chapters of Part 2 provide an up-to-date overview of current research in tactical portfolio construction and risk management. Their emphasis on general return distributions and tail risk measures is appropriate to the increasing penetration of hedge fund trading techniques into traditional fund and asset liability management. We hope that this treatment of tactical problems (and its companion strategic predecessor) will make a valuable contribution to the future practical use of systematic techniques in fund management. M.A.H. DEMPSTER, GAUTAM MITRA and GEORG C. PFLUG Cambridge, London & Vienna

PART

Dynamic Financial Planning

1

1

CHAPTER

Trends in Quantitative Equity Management: Survey Results FRANK J. FABOZZI, SERGIO FOCARDI and CAROLINE JONAS CONTENTS 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Growth in Equity Assets under Quantitative Management . 1.4 Changing Role for Models in Equity Portfolio Management 1.5 Modelling Methodologies and the Industry’s Evaluation . . . 1.6 Using High-Frequency Data . . . . . . . . . . . . . . . . . . . . . . 1.7 Measuring Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.8 Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.9 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.10 Looking Ahead . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.1

I

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

.3 .4 .5 .7 .8 11 11 12 13 14 15

INTRODUCTION

N THE SECOND HALF OF THE 1990s,

there was so much skepticism about quantitative fund management that Leinsweber (1999), a pioneer in applying advanced techniques borrowed from the world of physics to fund management, wrote an article entitled: ‘Is quantitative investment dead?’ In the article, Leinweber defended quantitative fund management and maintained that in an era of ever faster computers and ever larger databases, quantitative investment was here to stay. The skepticism towards quantitative fund management, provoked by the failure of some high-profile quantitative funds, was related to the fact that investment professionals felt that capturing market inefficiencies could best be done by exercising human judgement. Despite mainstream academic theory that had held that markets are efficient and unpredictable, the asset managers’ job has always been to capture market inefficiencies for their clients. At the academic level, the notion of efficient markets has been progressively relaxed. Empirical evidence that began to be accumulated in the 1970s led to the gradual 3

4 j CHAPTER 1

acceptance of the notion that financial markets are somewhat predictable and that systematic market inefficiencies can be detected (see Granger 1992 for a review to various models that accept departures from efficiency). Using the variance ratio test, Lo and MacKinlay (1988) disproved the random walk hypothesis. Additional insights return predictability was provided by Jegadeesh and Titman (1993), who established the existence of momentum phenomena. Since then, a growing number of studies have accumulated evidence that there are market anomalies that can be systematically exploited to earn excess profits after considering risk and transaction costs (see Pesaran 2005 for an up-todate presentation of the status of market efficiency). Lo (2004) proposed replacing the Efficient Market Hypothesis with the Adaptive Market Hypothesis arguing that market inefficiencies appear as the market adapts to changes in a competitive environment. The survey study described in this paper had as its objective to reveal to what extent the growing academic evidence that asset returns are predictable and that predictability can be exploited to earn a profit have impacted the way equity assets are being managed. Based on an Intertek 2003 survey on a somewhat different sample of firms, Fabozzi et al. (2004) revealed that models were used primarily for risk management, with many firms eschewing forecasting models. The 2006 survey reported in this chapter sought to reveal to what extent modelling has left the risk management domain to become full-fledged asset management methodology. Anticipating the results discussed below, the survey confirms that quantitative fund management is now an industrial reality, successfully competing with traditional asset managers for funds. Milevsky (2004) observes that the methods of quantitative finance have now been applied in the field of personal wealth management. We begin with a brief description of the field research methodology and the profile of responding firms. Section 1.3 discusses the central finding, that is, that models are being used to manage an increasing amount of equity asset value. Section 1.4 discusses the changing role of modelling in equity portfolio management, from decision-support systems to a fully automated portfolio construction and trading system, and from passive management to active management. Section 1.5 looks at the forecasting models most commonly used in the industry and discusses the industry’s evaluation of the techniques. Section 1.6 looks at the use (or lack of use) of high-frequency data and the motivating factors. Section 1.7 discusses risk measures being used and Section 1.8 optimization methodologies. The survey reveals a widespread use of optimization, which is behind the growing level of automation in fund management. The wide use of models has created a number of challenges: survey respondents say that differentiating quantitative products and improving on performance are a challenge. Lastly, in looking ahead, we discuss the issue of the role of models in market efficiency.

1.2

METHODOLOGY

The study is based on survey responses and conversations with industry representatives in 2006. In all, managers at 38 asset management firms managing a total of t3.3 trillion ($4.3 trillion) in equities participated in the survey. Participants include persons responsible for quantitative equity management and quantitative equity research at large and mediumsized firms in North America and Europe.

TRENDS IN QUANTITATIVE EQUITY MANAGEMENT

j

5

The home market of participating firms is 15 from North America (14 from U.S. 1 from Canada) and 23 from Europe (U.K. 7, Germany 5, Switzerland 4, Benelux 3, France 2 and Italy 2). Equities under management by participating firms range from t5 bn to t800 bn. While most firms whose use of quantitative methods is limited to performance analysis or risk measurement declined to participate in this study (only 5 of the 38 participating firms reported no equity funds under quantitative management), the study does reflect the use of quantitative methods in equity portfolio management at firms managing a total of t3.3 trillion ($4.3 trillion) in equities; 63) of the participating firms are among the largest asset managers in their respective countries. It is fair to say that these firms represent the way a large part of the industry is going with respect to the use of quantitative methods in equity portfolio management. (Note that of the 38 participants in this survey, 2 responded only partially to the questionnaire. For some questions, there are therefore 36 (not 38) responses.)

1.3

GROWTH IN EQUITY ASSETS UNDER QUANTITATIVE MANAGEMENT

The skepticism relative to the future of quantitative management at the end of the 1990s has given way and quantitative methods are now playing a large role in equity portfolio management. Twenty-nine percent (11/38) of the survey participants report that more than 75) of their equity assets are being managed quantitatively. This includes a wide spectrum of firms, with from t5 billion to over t500 billion in equity assets under management. Another 58) (22/38) report that they have some equities under quantitative management, though for most of these (15/22) the percentage of equities under quantitative management is less than 25)—often under 5)—of total equities under management. Thirteen percent (5/38) report no equities under quantitative management. Figure 1.1 represents the distribution of percentage of equities under quantitative management at different intervals for responding firms. Relative to the period 2004 2005, the amount of equities under quantitative management has grown at most firms participating in the survey. Eighty-four percent of the respondents (32/38) report that the percentage of equity assets under quantitative management has either increased with respect to 2004–2005 (25/38) or has remained stable at about 100) of equity assets (7/38). The percentage of equities under quantitative management was down at only one firm and stable at five. /

Number of Firms with Percentage of Equities under Quant Management in Different Intervals 11

More than 75% 50% to 74% 25% to 49%

2 5 15

More than 0 to 24% None

5

FIGURE 1.1 Distribution of the percentage of equities under quant management.

6 j CHAPTER 1 Contributing to a Wider Use of Quant Methods Other STP More/Better Data 3rd-Party s/w Desktop Computers Positive Results 0

20

40

60

80

100

120

FIGURE 1.2 Score attributed to each factor contributing to a wider use of quant methods.

One reason given by respondents to explain the growth in equity assets under quantitative management is the flows into existing quantitative funds. A source at a large U.S. asset management firm with more than 50) of its equities now under quantitative management said, ‘The firm has three distinct equity products: value, growth and quant. Quant is the biggest and is growing the fastest.’ The trend towards quantitative management is expected to continue. According to survey respondents, the most important factor contributing to a wider use of quantitative methods in equity portfolio management is the positive result obtained with these methods. Half of the participants rated positive results as the single most important factor contributing to the widespread use of quantitative methods. Other factors contributing to a wider use of quantitative methods in equity portfolio management are, in order of importance attributed to them by participants, the computational power now available on the desktop, more and better data, and the availability of third-party analytical software and visualization tools. Figure 1.2 represents the distribution of the score attributed to each factor. Participants were asked to rate from 1 to 5 in order of importance, 5 being the most important. Given the sample of 36 firms that responded, the maximum possible score is 180. Sources identified the prevailing in-house culture as the most important factor holding back a wider use of quantitative methods (this evaluation obviously does not hold for firms that can be described as quantitative): more than one third (10/27) of the respondents at other than quant-oriented firms considered this the major blocking factor. Figure 1.3 represents the distribution of the total score attributed to each factor. The positive evaluation of models in equity portfolio management is in contrast with the skepticism of some 10 years ago. A number of changes have occurred. First, expectations are now more realistic. In the 1980s and 1990s, traders were experimenting with methodologies from advanced science in hopes of making huge excess returns. Experience of the last 10 years has shown that models can indeed deliver but that their performance must be compatible with a well-functioning market.1 1 There was a performance decay in quantitatively managed equity funds in 2006 /2007. Many attribute this decaying performance to the fact that there are now more portfolio managers using the same factors and the same data.

TRENDS IN QUANTITATIVE EQUITY MANAGEMENT

j

7

Holding Back a Wider Use of Quant Methods Other Poor Results

2 37

Complexity

53

Cost Persons

52

Cost Data Cost IT In-House Culture

54 60 80

FIGURE 1.3 Score attributed to each factor holding back a wider use of quant methods.

Other technical reasons include a manifold increase in computing power and more and better data. Modellers have now available on their desktop computing power that, at the end of the 1980s, could be got only from multimillion dollar supercomputers. Data, including intraday data, can now be had (though the cost remains high) and are in general ‘cleaner’ and more complete. Current data include corporate actions, dividends, and fewer errors—at least in developed-country markets. In addition, investment firms (and institutional clients) have learned how to use models throughout the investment management process. Models are now part of an articulated process that, especially in the case of institutional investors, involves satisfying a number of different objectives, such as superior information ratios.

1.4

CHANGING ROLE FOR MODELS IN EQUITY PORTFOLIO MANAGEMENT

The survey reveals that quantitative models are now used in active management to find alphas (i.e. sources of excess returns), either relative to a benchmark or absolute. This is a considerable change with respect to the past when quantitative models were used primarily to manage risk and to select parsimonious portfolios for passive management. Another finding of this study is the growing amount of funds managed automatically by computer programs. The once futuristic vision of machines running funds automatically without the intervention of a portfolio manager is becoming a reality on a large scale: 55) of the respondents (21/38) report that at least part of their equity assets are now being managed automatically with quantitative methods; another three plan to automate at least a portion of their equity portfolios within the next 12 months. The growing automation of the equity investment process indicates that that there is no missing link in the technology chain that leads to automatic quantitative management. From return forecasting to portfolio formation and optimization, all the needed elements are in place. Until recently, optimization represented the missing technology link in the automation of portfolio engineering. Considered too brittle to be safely deployed, many firms eschewed optimization, limiting the use of modelling to stock ranking or risk control functions. Advances in robust estimation methodologies and in optimization now allow a manager to construct portfolios of hundreds of stocks chosen in universes of thousands of stocks with little or no human intervention outside of supervising the models.

8 j CHAPTER 1

1.5

MODELLING METHODOLOGIES AND THE INDUSTRY’S EVALUATION

At the end of the 1980s, academics and researchers at specialized quant boutiques experimented with many sophisticated modelling methodologies including chaos theory, fractals and multi-fractals, adaptive programming, learning theory, complexity theory, complex nonlinear stochastic models, data mining and artificial intelligence. Most of these efforts failed to live up to expectations. Perhaps expectations were too high. Or perhaps the resources or commitment required were lacking. Derman (2001) provides a lucid analysis of the difficulties that a quantitative analyst has to overcome. As observed by Derman, though modern quantitative finance uses some of the techniques of physics, a wide gap remains between the two disciplines. The modelling landscape revealed by the survey is simpler and more uniform. Regression analysis and momentum modelling are the most widely used techniques: respectively, 100) and 78) of the survey respondents say that these techniques are being used at their firms. Other modelling methods being widely used include cash flow analysis and behavioural modelling. Forty-seven percent (17/36) of the participating firms model cash flows; 44) (16/36) use behavioural modelling. Figure 1.4 represents the distribution of modelling methodologies among participants. Let us observe that regression models used today have undergone a substantial change since the first multifactor models such as Arbitrage Pricing Theory (APT) were introduced. Classical multifactor models such as APT are static models embodied in linear regression between returns and factors at the same time:

ri ¼ ai þ

p X

bij fj þ ei :

j¼1

Models of this type allow managers to measure risk but not to forecast returns, unless the factors are forecastable. Sources at traditional asset management firms typically use factor models to control risk or build stock screening systems. A source doing regression on factors to capture the risk-return trade-off of assets said, ‘Factor models are the most intuitive and most comprehensive models for explaining the sources of risk.’

Methodologies Used in Production Shrinkage/Averaging Regime Shifting Nonlinear Cointegration Cash Flow Behavioural Momentum/Reversal Regression

9 4 7 7 17 16 28

FIGURE 1.4 Distribution of modelling methodologies among participants.

36

TRENDS IN QUANTITATIVE EQUITY MANAGEMENT

j

9

However, modern regression models are dynamic models where returns at time t  1 are regressed on factors at time t: ri;tþ1 ¼ ai þ

p X

bij fj;t þ ei;t :

j¼1

Models of this type are forecasting models insofar as the factors at time t are predictors of returns at time behaviour t  1. In these models, individual return processes might exhibit zero autocorrelation but still be forecastable from other variables. Predictors might include financial and macroeconomic factors as well as company specific parameters such as financial ratios. Predictors might also include human judgment, for example analyst estimates, or technical factors that capture phenomena such as momentum. A source at a quant shop using regression to forecast returns said, ‘Regression on factors is the foundation of our model building. Ratios derived from financial statements serve as one of the most important components for predicting future stock returns. We use these ratios extensively in our bottom-up equity model and categorize them into five general categories: operating efficiency, financial strength, earnings quality (accruals), capital expenditures and external financing activities.’ Momentum and reversals are the second most widely used modelling technique among survey participants. In general, momentum and reversals are used as a strategy not as a model of asset returns. Momentum strategies are based on forming portfolios choosing the highest/lowest returns, where returns are estimated on specific time windows. Survey participants gave these strategies overall good marks but noted that (1) they do not always perform so well, (2) they can result in high turnover (though some use constraints/ penalties to deal with this problem) and (3) identifying the timing of reversals is tricky. Momentum was first reported in Jegadeesh and Titman (1993) in the U.S. market. Jegadeesh and Titman (2002) confirm that momentum continued to exist in the 1990s in the US market throughout the 1990s. Karolyi and Kho (2004) examined different models for explaining momentum and introduced a new bootstrap test. Karolyi and Kho conclude that no random walk or autoregressive model is able to explain the magnitude of momentum empirically found; they suggest that models with time varying expected returns come closer to explaining the empirical magnitude of momentum. Momentum and reversals are presently explained in the context of local models updated in real time. For example, momentum as described in Jegadeesh and Titman (1993) is based on the fact that stock prices can be represented as independent random walks when considering periods of the length of one year. However, it is fair to say that there is no complete agreement on the econometrics of asset returns that would justify momentum and reversals and stylized facts on a global scale, and not as local models. It would be beneficial to know more about the econometrics of asset returns that sustain momentum and reversals. Behavioural phenomena are considered to play an important role in asset predictability; as mentioned, 44) of the survey respondents say they use behavioural modelling. Behavioural modellers attempt to capture phenomena such as departures from rationality

10 j CHAPTER 1

on the part of investors (e.g. belief persistence), patterns in analyst estimates, and corporate executive investment/disinvestment behaviour. Behavioural finance is related to momentum in that the latter is often attributed to various phenomena of persistence in analyst estimates and investor perceptions. A source at a large investment firm that has incorporated behavioural modelling into its active equity strategies commented, ‘The attraction of behavioural finance is now much stronger than it was just five years ago. Everyone now acknowledges that markets are not efficient, that there are behavioural anomalies. In the past, there was the theory that was saying that markets are efficient while market participants such as the proprietary trading desks ignored the theory and tried to profit from the anomalies. We are now seeing a fusion of theory and practice.’ We remark that the term behavioural modelling is often used rather loosely. Fullfledged behavioural modelling exploits a knowledge of human psychology to identify situations where investors are prone to show behaviour that leads to market inefficiencies. The tendency now is to call ‘behavioural’ any model that exploits market inefficiency. However, implementing true behavioural modelling is a serious challenge. Even firms with very large, powerful quant teams say that ‘considerable work is required to translate [departures from rationality] into a set of rules for identifying stocks as well as entry and exit points for a quantitative stock selection process.’ As for other methodologies used in return forecasting, sources cited nonlinear methods and co-integration. Nonlinear methods are being used to model return processes at 19) (7/36) of the responding firms. The nonlinear method most widely used among survey participants is classification and regression trees (CART). The advantage of CART is its simplicity and the ability of CART methods to be cast in an intuitive framework. A source using CART as a central part of the portfolio construction process in enhanced index and longer-term value-based portfolios said, ‘CART compresses a large volume of data into a form which identifies its essential characteristics, so the output is easy to understand. CART is non-parametric—which means that it can handle an infinitely wide range of statistical distributions—and nonlinear so as a variable selection technique it is particularly good at handling higher-order interactions between variables.’ Only 11) (4/36) of the respondents use nonlinear regime-shifting models; at most firms, judgment is used to assess regime change. Obstacles to modelling regime shifts include the difficulty in detecting the precise timing of a regime switch and the very long time series required for true estimation. A source at a firm where regime-shifting models have been experimented with commented, ‘Everyone knows that returns are conditioned by market regimes, but the potential for overfitting when implementing regime-switching models is great. If you could go back with fifty years of data—but we have only some ten years of data and this is not enough to build a decent model.’ Co-integration is being used by 19) (7/36) of the respondents. Co-integration models the short-term dynamics (direction) and long-run equilibrium (fair value). A perceived plus of co-integration is the transparency that it provides: the models are based on economic and finance theory and calculated from economic data.

TRENDS IN QUANTITATIVE EQUITY MANAGEMENT

1.6

j

11

USING HIGH-FREQUENCY DATA

High frequency data (HFD) are being used at only 14) of the responding firms (5/36), to identify profit opportunities and improve forecasts. Another three plan to use HFD within the next 12 months. A source at a large investment firm that is using HFD said, ‘We use high-frequency data in event studies. The objective is to gain an understanding of the mechanisms of the market.’ A source which is planning to use high-frequency data in the coming 12 months remarked, ‘We believe that high-frequency data will allow us to evaluate exactly when it is optimal to trade, for example at close, VWAP, or midday, and to monitor potential market impact of our trades and potential front-running of our brokers.’ (VWAP stands for volume-weighted average price.) Though it is believed that HFD could be useful, cost of data is the blocking factor. Survey participants voiced concerns that the cost of data will hamper the development of models in the future. One source observes, ‘The quasi monopolistic positioning of data vendors allows them to charge prices that are incompatible with the revenues structure of all but the biggest firms.’ Other reasons cited by the sources not using HFD are a (perceived) unattractive noise-to-signal ratio and resistance to HFD-based strategies on the part of institutional investors.

1.7

MEASURING RISK

Risk is being measured at all the responding firms. Risk measures most widely used among participants include variance (97) or 35/36), Value at Risk (VaR) (67) or 24/36) and downside risk measures (39) or 14/36), Conditional VaR (CVaR), and extreme value theory (EVT) are used at 4 (11)) and 2 (6)) firms, respectively. The considerable use of asymmetric risk measures such as downside risk can be ascribed to the growing popularity of financial products with guaranteed returns. Many firms compute several risk measures: the challenge here is to merge the different risk views into a coherent risk assessment. Figure 1.5 represents the distribution of risk measures used by participants. It is also interesting to note that among survey participants, there is a heightened attention to model risk. Model averaging and shrinkage techniques are being used by onefourth (9/36) of the survey participants. The recent take-up of these techniques is related to the fact that most firms are now using multiple models to forecast returns, a trend that is up compared to two or three years ago. Other techniques to mitigate model risk, such as Risk Measures Being Used EVT

2 14

Downside Risk CVaR VaR

4 24

Variance

FIGURE 1.5 Distribution of risk measures adopted by participants.

35

12 j CHAPTER 1

random coefficient models, are not used much in the industry. In dealing with model risk we must distinguish between averaging model results and averaging models themselves. The latter technique, embodied in random coefficient models, is more difficult and requires more data.

1.8

OPTIMIZATION

Another area where much has changed recently is optimization. According to sources, optimization is now performed at 92) (33/36) of the participating firms, albeit in some cases only rarely. Mean-variance is the most widely used technique among survey participants: it is being used by 83) (30/36) of the respondents. It is followed by utility optimization (42) or 15/36) and, more recently, robust optimization (25) or 9/36). Only one firm mentioned that it is using stochastic optimization. Figure 1.6 represents the distribution of optimization methods. The wider use of optimization is a significant development compared to just a few years ago when many sources reported that they eschewed optimization: the difficulty of identifying the forecasting error was behind the then widely held opinion that optimization techniques were too brittle and prone to ‘error maximization.’ The greater use of optimization is due to advances in large-scale optimization coupled with the ability to include constraints and robust methods for estimation and optimization itself. It is significant: portfolio formation strategies rely on optimization. With optimization now feasible, the door is open to a fully automated investment process. In this context, it is noteworthy that 55) of the survey respondents report that at least a portion of their equity assets is being managed by a fully automated process. Optimization is the engineering part of portfolio construction. Most portfolio construction problems can be cast in an optimization framework, where optimization is applied to obtain the desired optimal risk-return profile. Optimization is the technology behind the current offering of products with specially engineered returns, such as guaranteed returns. However, the offering of products with particular risk-return profiles requires optimization methodologies that go well beyond the classical mean-variance optimization. In particular one must be able to (1) work with real-world utility functions and (2) apply constraints to the optimization process.

Optimization Methods in Use Stochastic Opt

1 9

Robust Opt

15

Utility Opt

30

Mean-Var None

3

FIGURE 1.6 Distribution of optimization methods adopted by participants.

TRENDS IN QUANTITATIVE EQUITY MANAGEMENT

1.9

j

13

CHALLENGES

The growing diffusion of models is not without challenges. Survey participants noted three: “ “ “

increasing difficulty in differentiating products; difficulty in marketing quant funds, especially to non-institutional investors; and performance decay.

Quantitative equity management has now become so widespread that a source at a long-established quantitative investment firm remarked, ‘There is now a lot of competition from new firms entering the space [of quantitative investment management]. The challenge is to continue to distinguish ourselves from competition in the minds of clients.’ With many quantitative funds based on the same methodologies and using the same data, the risk is to construct products with the same risk-return profile. The head of active equities at a large quantitative firm with more than a decade of experience in quantitative management remarked, ‘Everyone is using the same data and reading the same articles: it’s tough to differentiate.’ While sources report that client demand is behind the growth of (new) pure quantitative funds, some mentioned that quantitative funds might be something of a hard sell. A source at a medium-sized asset management firm servicing both institutional clients and high net worth individuals said, ‘Though clearly the trend towards quantitative funds is up, quant approaches remain difficult to sell to private clients: they remain too complex to explain, there are too few stories to tell, and they often have low alpha. Private clients do not care about high information ratios.’ Markets are also affecting the performance of quantitative strategies. A recently released report from the Bank for International Settlements (2006) noted that this is a period of historically low volatility. What is exceptional about this period, observes the report, is the simultaneous drop in volatility in all variables: stock returns, bond spread, rates and so on. While the role of models in reducing volatility is unclear, what is clear is that models produce a rather uniform behaviour. Quantitative funds try to differentiate themselves either finding new unexploited sources of return forecastability, for example novel ways of looking at financial statements, or using optimization creatively to engineer special riskreturn profiles. A potentially more serious problem is performance decay. Survey participants remarked that model performance was not so stable. Firms are tackling these problems in two ways. First, they are protecting themselves from model breakdown with model risk mitigation techniques, namely by averaging results obtained with different models. It is unlikely that all models breakdown in the same way in the same moment, so that averaging with different models allows managers to diversify model risk. Second, there is an on-going quest for new factors, new predictors, and new aggregations of factors and predictors. In the long run, however, something more substantial might be required: this is the subject of the next and final section.

14 j CHAPTER 1

1.10

LOOKING AHEAD

Looking ahead, we can see a number of additional challenges. Robust optimization, robust estimation and the integration of the two are probably on the research agenda of many firms. As asset management firms strive to propose innovative products, robust and flexible optimization methods will be high on the R & D agenda. In addition, as asset management firms try to offer investment strategies to meet a stream of liabilities (i.e., measured against liability benchmarking), multistage stochastic optimization methods will become a priority for firms wanting to compete in this arena. Pan et al. (2006) call ‘Intelligent Finance’ the new field of theoretical finance at the confluence of different scientific disciplines. According to the authors, the theoretical framework of intelligent finance consists of four major components: financial information fusion, multilevel stochastic dynamic process models, active portfolio and total risk management, and financial strategic analysis. The future role of high-frequency data is not yet clear. HFD are being used (1) to improve on model quality thanks to the 2000-fold increase in sample size they offer with respect to daily data and (2) to find intraday profit opportunities. The ability to improve on model quality thanks to HFD is the subject of research. It is already known that quantities such as volatility can be measured with higher precision using HFD. Using HFD, volatility ceases to be a hidden variable and becomes the measurable realized volatility, introduced by Torbin et al. (2003). If, and how, this increased accuracy impacts models whose time horizon is in the order of weeks or months is a subject not entirely explored. It might be that in modelling HFD one captures short-term effects that disappear at longer time horizons. Regardless of the frequency of data sampling, modellers have to face the problem of performance decay that is the consequence of a wider use of models. Classical financial theory assumes that agents are perfect forecasters in the sense that they know the stochastic processes of prices and returns. Agents do not make systematic predictable mistakes: their action keeps the market efficient. This is the basic idea underlying rational expectations and the intertemporal models of Merton. Practitioners (and now also academics) have relaxed the hypothesis of the universal validity of market efficiency; indeed, practitioners have always being looking for asset mispricings that could produce alpha. As we have seen, it is widely believed that mispricings are due to behavioural phenomena, such as belief persistence. This behaviour creates biases in agent evaluations—biases that models attempt to exploit in applications such as momentum strategies. However, the action of models tends to destroy the same sources of profit that they are trying to exploit. This fact receives attention in applications such as measuring the impact of trades. In almost all current implementations, measuring the impact of trades means measuring the speed at which models constrain markets to return to an unprofitable efficiency. To our knowledge, no market impact model attempts to measure the opposite effect, that is, the eventual momentum induced by a trade. It is reasonable to assume that the diffusion of models will reduce the mispricings due to behavioural phenomena. However, one might reasonably ask whether the action of

TRENDS IN QUANTITATIVE EQUITY MANAGEMENT

j

15

models will ultimately make markets more efficient, destroying any residual profitability in excess of market returns, or if the action of models will create new opportunities that can be exploited by other models, eventually by a new generation of models based on an accurate analysis of model biases. It is far from being obvious that markets populated by agents embodied in mathematical models will move toward efficiency. In fact, models might create biases of their own. For example, momentum strategies (buy winners, sell losers) are a catalyst for increased momentum, farther increasing the price of winners and depressing the price of losers. This subject has received much attention in the past as researchers studied the behaviour of markets populated by boundedly rational agents. While it is basically impossible, or at least impractical, to code the behaviour of human agents, models belong to a number of well-defined categories that process past data to form forecasts. Studies, based either on theory or on simulation, have attempted to analyse the behaviour of markets populated by agents that have bounded rationality, that is, filter past data to form forecasts.2 One challenge going forward will be to understand what type of inefficiencies are produced by markets populated by automatic decision-makers whose decisions are based on past data. It is foreseeable that simulation and artificial markets will play a greater role as discovery devices.

REFERENCES Bank for International Settlements, The recent behaviour of financial market volatility. BIS Release Paper, 29 August, 2006. Derman, E., A guide for the perplexed quant. Quant. Finan., 2001, 1(5), 476 /480. Fabozzi, F.J., Focardi, S.M. and Jonas, C.L., Trends in quantitative asset management in Europe. J. Portfolio Manag., 2004, 31(4), 125/132 (Special European Section). Granger, C.W.J., Forecasting stock market prices: Lessons for forecasters. Int. J. Forecast., 1992, 8, 3 /13. Jegadeesh, N. and Titman, S., Returns to buying winners and selling losers: Implications for stock market efficiency. J. Finan., 1993, 48(1), 65 /92. Jegadeesh, N. and Titman, S., Cross-sectional and time-series determinants of momentum returns. Rev. Financ. Stud., 2002, 15(1), 143/158. Kahneman, D., Maps of bounded rationality: Psychology for behavioral economics. Am. Econ. Rev., 2003, 93(5), 1449/1475. Karolyi, G.A. and Kho, B.-C., Momentum strategies: Some bootstrap tests. J. Empiric. Finan., 2004, 11, 509/536. LeBaron, B., Agent-based computational finance. In Handbook of Computational Economics, edited by L. Tesfatsion and K. Judd, 2006 (North-Holland: Amsterdam). Leinweber, D., Is quantitative investing dead? Pensions & Investments, 8 February, 1999. Lo, A., The adaptive markets hypothesis. J. Portfolio Manag., 2004, 30 (Anniversary Issue), 15 /29. Lo, A. and MacKinley, C., Stock market prices do not follow random walks: Evidence from a simple specification test. Rev. Financ. Stud., 1988, 1, 41 /66. Milevsky, M., A diffusive wander through human life. Quant. Finan., 2004, 4(2), 21 /23. 2

See Sargent (1994) and Kahneman (2003) for the the theoretical underpinning of bounded rationality from the statistical and behavioural point of view, respectively. See LeBaron (2006) for a survey of research on computational finance with boundedly rational agents.

16 j CHAPTER 1

Pan, H., Sornette, D. and Kortanek, K., Intelligent finance—an emerging direction. Quant. Finan., 2006, 6(4), 273/277. Pesaran, M.H., Market efficiency today. Working Paper 05.41, 2005 (Institute of Economic Policy Research). Sargent, J.T., Bounded Rationality in Macroeconomics, 1994 (Oxford University Press: New York). Torbin, A., Bollerslev, T., Diebold, F.X. and Labys, P., Modeling and forecasting realized volatility. Econometrica, 2003, 71(2), 579 /625.

CHAPTER

2

Portfolio Optimization under the Value-at-Risk Constraint TRAIAN A. PIRVU CONTENTS 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.1.1 Our Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .19 2.2 Model Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.2.1 The Financial Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .19 2.2.2 Consumption, Trading Strategies and Wealth. . . . . . . . . . . . . . . . . .20 2.2.3 Value-at-Risk Limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .21 2.3 Objective. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.4 Logarithmic Utility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.5 Nonlogarithmic Utility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.6 Numerical Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2.7 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 Appendix 2.A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

2.1

M

INTRODUCTION

of their traders by imposing limits on the risk of their portfolios. Lately, the Value-at-Risk (VaR) risk measure has become a tool used to accomplish this purpose. The increased popularity of this risk measure is due to the fact that VaR is easily understood. It is the maximum loss of a portfolio over a given horizon, at a given confidence level. The Basle Committee on Banking Supervision requires U.S. banks to use VaR in determining the minimum capital required for their trading portfolios. In the following we give a brief description of the existing literature. Basak and Shapiro (2001) analyse the optimal dynamic portfolio and wealth-consumption policies of utility maximizing investors who must manage risk exposure using VaR. They find that VaR risk ANAGERS LIMIT THE RISKINESS

17

18 j CHAPTER 2

managers pick a larger exposure to risky assets than non-risk managers, and consequently incur larger losses when losses occur. In order to fix this deficiency they choose another risk measure based on the risk-neutral expectation of a loss. They call this risk measure Limited Expected Loss (LEL). One drawback of their model is that the portfolios VaR is never re-evaluated after the initial date, making the problem a static one. In a similar setup, Berkelaar et al. (2002) show that, in equilibrium, VaR reduces market volatility, but in some cases raises the probability of extreme losses. Emmer et al. (2001) consider a dynamic model with Capital-at-Risk (a version of VaR) limits. However, they assume that portfolio proportions are held fixed during the whole investment horizon and thus the problem becomes a static one as well. Cuoco et al. (2001) develop a more realistic dynamically consistent model of the optimal behaviour of a trader subject to risk constraints. They assume that the risk of the trading portfolio is re-evaluated dynamically by using the conditioning information, and hence the trader must satisfy the risk limit continuously. Another assumption they make is that when assessing the risk of a portfolio, the proportions of different assets held in the portfolio are kept unchanged. Besides the VaR risk measure, they consider a coherent risk measure Tail Value at Risk (TVaR), and establish that it is possible to identify a dynamic VaR risk limit equivalent to a dynamic TVaR limit. Another of their findings is that the risk exposure of a trader subject to VaR and TVaR limits is always lower than that of an unconstrained trader. Pirvu (2005) started with the model of Cuoco et al. (2001). We find the optimal growth portfolio subject to these risk measures. The main finding is that the optimal policy is a projection of the optimal portfolio of an unconstrained log agent (the Merton proportion) onto the constraint set with respect to the inner product induced by the volatility matrix of the risky assets. Closed-form solutions are derived even when the constraint set depends on the current wealth level. Cuoco and Liu (2003) study the dynamic investment and reporting problem of a financial institution subject to capital requirements based on self-reported VaR estimates. They show that optimal portfolios display a local three-fund property. Leippold et al. (2002) analyse VaR-based regulation rules and their possible distortion effects on financial markets. In partial equilibrium the effectiveness of VaR regulation is closely linked to the leverage effect, the tendency of volatility to increase when the prices decline. Vidovic et al. (2003) considered a model with time-dependent parameters, but the risk constraints were imposed in a static fashion. Yiu (2004) looks at the optimal portfolio problem, when an economic agent is maximizing the utility of her intertemporal consumption over a period of time under a dynamic VaR constraint. A numerical method is proposed to solve the corresponding HJB equation. They find that investment in risky assets is optimally reduced by the VaR constraint. Atkinson and Papakokinou (2005) derive the solution to the optimal portfolio and consumption subject to CaR and VaR constraints using stochastic dynamic programming. This paper extends Pirvu (2005) by allowing for intertemporal consumption. We address an issue raised by Yiu (2004) and Atkinson and Papakokinou (2005) by considering a

PORTFOLIO OPTIMIZATION UNDER THE VALUE-AT-RISK CONSTRAINT

j

19

market with random coefficients. It was also suggested as a new research direction by Cuoco et al. (2001). To the best of our knowledge this is the first work in portfolio choice theory with CRRA-type preferences, time-dependent market coefficients, incomplete financial markets, and dynamically consistent risk constraints in a Brownian motion framework. 2.1.1

Our Contribution

We propose a new approach with the potential for numerical applications. The main idea is to consider, on every probabilistic path, an auxiliary deterministic control problem, which we analyse. The existence of a solution of the deterministic control problem does not follow from classical results. We establish it and we also show that first-order necessary conditions are also sufficient for optimality. We prove that a solution of this deterministic control problem is the optimal portfolio policy for a given path. The advantage of our method over classical methods is that it allows for a better numerical treatment. The remainder of this chapter is organized as follows. Section 2.2 describes the model, including the definition of the Value-at-Risk constraint. Section 2.3 formulates the objective and shows the limitations of the stochastic dynamic programming approach in this context. Section 2.4 treats the special case of logarithmic utility. The problem of maximizing expected logarithmic utility of intertemporal consumption is solved in closed form. This is done by reducing it to a nonlinear program, which is solved pathwise. One finding is that, at the final time, the agent invests the least proportion of her wealth in stocks. The optimal policy is a projection of the optimal portfolio and consumption of an unconstrained agent onto the constraint set. Section 2.5 treats the case of power utility, in the totally unhedgable market coefficients paradigm (see Example 2.7.4, p. 305 of Karatzas and Shreve 1998). The stochastic control portfolio problem is transformed into a deterministic control problem. The solution is characterized by the Pontryagin maximum principle (first-order necessary conditions). Section 2.6 contains an appropriate discretization of the deterministic control problem. It leads to a nonlinear program that can be solved by standard methods. It turns out that the necessary conditions of the discretized problem converge to the necessary conditions of the continuous problem. For numerical experiments, one can use NPSOL, a software package for solving constrained optimization problems that employs a sequential quadratic programming (SQP) algorithm. We end this section with some numerical experiments. The conclusions are summarized in Section 2.7. We conclude the paper with an appendix containing the proofs of the lemmas.

2.2

MODEL DESCRIPTION

2.2.1

The Financial Market

Our model of a financial market, based on a filtered probability space ðO; fF t g0t1 ; F ; PÞ satisfying the usual conditions, consists of m
1 assets. The first, fS0 ðtÞgt2½0;1 , is a riskless bond with a positive interest rate r, i.e. dS0 ðtÞ ¼ S0 ðtÞr dt. The remaining m are stocks and evolve according to the following stochastic differential equation:

20 j CHAPTER 2

" dSi ðtÞ ¼ Si ðtÞ ai ðtÞdt þ

n X

# sij ðtÞdWj ðtÞ ;

j¼1

ð2:1Þ

0  t  1; i ¼ 1; . . . ; m;

where the process fW ðtÞgt2½0;1Þ ¼ fðWj ðtÞÞj¼1;...;n gt2½0;1Þ is an n-dimensional standard Brownian motion. Here, faðtÞgt2½0;1Þ ¼ fðai ðtÞÞi¼1;...;m gt2½0;1Þ is an Rm -valued mean rate j¼1;...;n of return process, and fsðtÞgt2½0;1Þ ¼ fðsij ðtÞÞi¼1;...;m gt2½0;1Þ is an m n matrix-valued volatility process. We impose the following regularity assumptions on the coefficient processes a(t) and s(t).





All the components of the process faðtÞgt2½0;1Þ are assumed positive, continuous and fF t g-adapted. The matrix-valued volatility process fsðtÞgt2½0;1Þ is assumed continuous, fF t gadapted and with linearly independent rows for all t 2 ½0; 1Þ, a.s.

The last assumption precludes the existence of a redundant asset and arbitrage opportunities. The rate of excess return is the Rm -valued process fmðtÞgt2½0;1Þ ¼ fðmi ðtÞÞi¼1;...;m gt2½0;1Þ , with mi ðtÞ ¼ ai ðtÞ r, which is assumed positive. This also covers the case of an incomplete market if n
m (more sources of randomness than stocks). 2.2.2

Consumption, Trading Strategies and Wealth

In this model the agent is allowed to consume. The intermediate consumption process, denoted fCðtÞgt2½0;1Þ , is assumed positive, and fF t g-progressively measurable. Let fðzðtÞ; cðtÞÞgt2½0;1Þ ¼ fðzi ðtÞi¼1;...;m ; cðtÞgt2½0;1Þ be an Rmþ1 -valued portfolio-proportion process. At time t its components are the proportions of the agent’s wealth invested in stocks, z(t), and her consumption rate, c(t). An Rmþ1 -valued portfolio-proportion process is called admissible if it is fF t g-progressively measurable and satisfies Z

t

 zT ðuÞmðuÞdu þ

0

þ

Z

t

 zT ðuÞsðuÞ2 du

0

Z

ð2:2Þ

t

cðuÞdu < 1; a.s.;

8t 2 ½0; 1Þ;

0

where, as usual, I ×I is the standard Euclidean norm in Rm . Given fðzðtÞ; cðtÞÞgt2½0;1Þ is a Pm portfolio-proportion process, the leftover wealth X z;c ðtÞð1 i¼1 zi ðtÞÞ is invested in the riskless bond S0(t). It may be that this quantity is negative, in which case we are borrowing at rate r
0. The dynamics of the wealth process fX z;c ðtÞgt2½0;1Þ of an agent using the portfolio-proportion process fðzðtÞ; cðtÞÞgt2½0;1Þ is given by the following stochastic differential equation:

PORTFOLIO OPTIMIZATION UNDER THE VALUE-AT-RISK CONSTRAINT

j

21

 dX z;c ðtÞ ¼ X z;c ðtÞ zT ðtÞaðtÞ cðtÞ dt þ zT ðtÞsðtÞdW ðtÞ ! m X zi ðtÞ X z;c ðtÞr dt þ 1 i¼1

 ¼ X z;c ðtÞ r cðtÞ þ zT ðtÞmðtÞ dt þ zT ðtÞsðtÞdW ðtÞÞ: Let us define the p-quadratic correction to the saving rate r: 4

Qp ðt; z; cÞ ¼ r c þ zT mðtÞ þ

 p 1 zT sðtÞ2 : 2

ð2:3Þ

The above stochastic differential equation has a unique strong solution if (2.2) is satisfied and is given by the explicit expression z;c

X ðtÞ ¼ Xð0Þ exp

Z 0

t

Q0 ðu; zðuÞ; cðuÞÞdu þ

Z

t T

z ðuÞsðuÞdW ðuÞ :

ð2:4Þ

0

The initial wealth X z;c ð0Þ ¼ Xð0Þ takes values in (0,) and is exogenously given. 2.2.3

Value-at-Risk Limits

For the purposes of risk measurement, one can use an approximation of the distribution of the investor’s wealth at a future date. A detailed explanation of why this practice should be employed can be found on p. 8 of Cuoco et al. (2001) (see also p. 18 of Leippold et al. (2002)). Given a fixed time instance t ] 0, and a length t
0 of the measurement horizon [t, t
t], the projected distribution of wealth from trading and consumption is usually calculated under the assumptions that 1. the portfolio proportion process fðzðuÞ; cðuÞÞgu2½t;tþt , as well as 2. the market coefficients fðaðuÞgu2½t;tþt Þ and fðsðuÞgu2½t;tþt , will stay constant and equal their present value throughout [t, t
t]. If t is small, for example t  1 trading day, the market coefficients will not change much and this supports assumption 2. The wealth’s dynamics equation yields the projected wealth at t
t:  X z;c ðt þ tÞ ¼ X z;c ðtÞ exp Q0 ðt; zðtÞ; cðtÞÞt

 þ zT ðtÞsðtÞðW ðt þ tÞ W ðtÞÞ ;

whence the projected wealth loss on the time interval [t, t
t] is  X z;c ðtÞ X z;c ðt þ tÞ ¼ X z;c ðtÞ½1 exp Q0 ðt; zðtÞ; cðtÞÞt  þ zT ðtÞsðtÞðW ðt þ tÞ W ðtÞÞ :

22 j CHAPTER 2

The random variable zT ðtÞsðtÞðW ðt þ tÞ W ðtÞÞ is, conditionally on F t , normally pffiffiffi distributed with mean zero and standard deviation kzT ðtÞsðtÞk t. Let the confidence parameter a 2 ð0; 1=2 be exogenously specified. The a-percentile of the projected loss X z;c ðtÞ X z;c ðt þ tÞ conditionally on F t is   pffiffiffi X z;c ðtÞ 1 exp Q0 ðt; zðtÞ; cðtÞÞt þ N 1 ðaÞkzT ðtÞsðtÞk t ; where N( ×) denotes the standard cumulative normal distribution function. This prompts the Value-at-Risk (VaR) of projected loss  pffiffiffiþ  4  : VaRðt; z; c; xÞ ¼ x 1 exp Q0 ðt; z; cÞt þ N 1 ðaÞzT sðtÞ t Let aV  (0,1) be an exogenous risk limit. The Value-at-Risk constraint is that the agent at every time instant t ] 0 must choose a portfolio proportion (z(t), c(t)) that would result in a relative VaR of the projected loss on [t, t
t] less than aV . This, strictly speaking, is the set of all admissible portfolios which, for all t ] 0, belong to FV(t), defined by

VaRðt; z; c; xÞ  aV : FV ðtÞ ¼ ðz; cÞ 2 R  ½0; 1Þ; x 4



m

ð2:5Þ

The fraction VaRux rather than VaR is employed, whence the name relative VaR. If one imposes VaR in absolute terms, the constraint set depends on the current wealth level and this makes the analysis more involved (see Cuoco et al. 2001; Pirvu 2005). For a given path v let us denote oðtÞ ¼ ðos Þst as the projection up to time t of its trajectory. One can see that, for a fixed vt, the set FV(t) is compact and convex, being the level set of a convex, unbounded function fV(t, z, c), ( FV ðtÞ ¼

) 1 ; ðz; cÞ 2 R  ½0; 1Þ; fV ðt; z; cÞ  log 1 aV m

where  pffiffiffi 4 fV ðt; z; cÞ ¼ Q0 ðt; z; cÞt N 1 ðaÞzT sðtÞ t:

ð2:6Þ

The function fV , although quadratic in z and linear in c, may still fail to be convex in (z, c) if a ]1u2, thus FV (t) may not be a convex set (see Figure 2.1, p. 10 of Cuoco et al. 2001). However, the choice of a 2 ð0; ð1=2Þ makes N 1 ðaÞ  0 and this yields convexity.

2.3

OBJECTIVE

Let finite time horizon T and the discount factor d (the agent’s impatient factor) be primitives of the model. Given X(0), the agent seeks to choose an admissible portfolioproportion process such that ðzðtÞ; cðtÞÞ 2 FV ðtÞ for all 0 5t 5T, and the expected value of her CRRA utility of intertemporal consumption and final wealth,

PORTFOLIO OPTIMIZATION UNDER THE VALUE-AT-RISK CONSTRAINT

Z

j

23

T

e dt Up ðCðtÞÞdt þ e dT Up ðX z;c ðT ÞÞ;

ð2:7Þ

0

is maximized over all admissible portfolios processes satisfying the same constraint. Here, 4 Up ðxÞ ¼ x p =p, with p B1 the coefficient of relative risk aversion (CRRA). Let us assume for the moment that the market coefficients are constants. In this case the constraint set FV(t) does not change over time and we denote it by FV . Then one can use dynamic programming techniques to characterize the optimal portfolio and consumption policy. The problem is to find a solution to the HJB equation. Define the optimal value function as

V ðx; tÞ ¼ max Et ðz;cÞ2FV

Z

T

e

dt

Up ðCðuÞÞdu þ e

dT

  z;c Up X ðT Þ ;

t

where Et is the conditional operator given the information known up to time t and X z;c ðtÞ ¼ x. The HJB equation is max J ðt; x; z; cÞ ¼ 0;

ðz;cÞ2FV

where 4

 @V @V þ x r c þ zT m @t @x 2 kzT sk x 2 @ 2 V þ ; @x 2 2

J ðt; x; c; zÞ ¼ e dt Up ðcxÞ þ

with the boundary condition V ðx; T Þ ¼ e dT Up ðxÞ. The value function V inherits the concavity of the utility functions Up. Being jointly concave in (z, c), the function J is maximized over the set FV at a unique point ð z; cÞ. Moreover, this point should lie on the boundary of FV and one can derive first-order optimality conditions by means of Lagrange multipliers. Together with the HJB equation this yields a highly nonlinear PDE that is hard to solve numerically (a numerical scheme was proposed by Yiu (2004), but no convergence result was reported). In the following we approach the portfolio optimization problem by reducing it to a deterministic control problem. We are then able to obtain explicit solutions for logarithmic utility.

2.4

LOGARITHMIC UTILITY

Let us consider the case where the agent is deriving utility from intermediate consumption only. It is straightforward to extend it to also encompass the utility of the final wealth. In light of (2.4),

24 j CHAPTER 2

Z

z;c

t

log X ðtÞ ¼ log Xð0Þ þ Q0 ðs; zðsÞ; cðsÞÞds 0 Z t zT ðsÞsðsÞdW ðsÞ: þ

ð2:8Þ

0

What facilitates the analysis is the decomposition of the utility from intertemporal consumption into a signal, a Lebesque integral and noise, which comes at an Itoˆ integral rate. The decomposition is additive and the expectation operator cancels the noise. Indeed, Z

T

e

dt

Z

T

e dt log ðcðtÞX z;c ðtÞÞdt 0 Z T 1 e dT log Xð0Þ þ ¼ e dt log cðtÞdt d 0 Z T Z t þ e dt Q0 ðs; zðsÞ; cðsÞÞds dt 0 0 Z t Z T

dt þ e zT ðsÞsðsÞdW ðsÞdt:

log CðtÞdt ¼

0

0

0

By Fubini’s theorem Z

T

Z

t

e 0

dt

Q0 ðs; zðsÞ; cðsÞÞds dt ¼

0

Z T Z

e 0

¼

Z

0

T

e



T

s

dt

dt

Q0 ðs; zðsÞ; cðsÞÞdt ds

e dT d

Q0 ðt; zðtÞ; cðtÞÞdt;

hence Z 0

T

e dt log CðtÞdt ¼

1 e dT

log Xð0Þ   1

dt

dðT tÞ þ e log cðtÞ þ ð1 e ÞQ0 ðt; zðtÞ; cðtÞÞ dt d 0 Z t Z T

dt e zT ðsÞsðsÞdW ðsÞdt: þ d Z T

0

ð2:9Þ

0

The linearly independent rows assumption on the matrix-valued volatility process yields the existence of the inverse ðsðtÞsT ðtÞÞ 1 and so equation sðtÞsT ðtÞzM ðtÞ ¼ mðtÞ

ð2:10Þ

1

uniquely defines the stochastic process fzM ðtÞgt2½0;1Þ ¼ fðsðtÞðsT ðtÞÞ mðtÞgt2½0;1Þ , called the Merton-ratio process. It has the property that it maximizes (in the absence of portfolio constraints), the rate of growth, and the log-optimizing investor would invest exactly using

PORTFOLIO OPTIMIZATION UNDER THE VALUE-AT-RISK CONSTRAINT

j

25

the components of zM(t) as portfolios proportions (see Section 3.10 of Karatzas and Shreve 1991). By (2.10)  T  z ðtÞsðtÞ2 ¼ zT ðtÞmðtÞ ¼ mT ðtÞðsðtÞsT ðtÞÞ 1 mðtÞ: M

M

ð2:11Þ

The following integrability assumption is rather technical, but it guarantees that a local martingale (Itoˆ integral) is a (true) martingale (see p. 130 of Karatzas and Shreve 1991). Let us assume that Z

T

 zT ðuÞsðuÞ2 du < 1:

E

M

ð2:12Þ

0

This requirement, although imposed on the market coefficients (see Equation (2.11)), is also inherited for all portfolios satisfying the Value-at-Risk constraint. Lemma 2.1: For every R t ðzðtÞ; cðtÞÞ 2 FV ðtÞ the process martingale, hence E 0 zT ðsÞsðsÞ dW ðsÞ ¼ 0.

Rt 0

zTs ðsÞsðsÞ dW ðsÞ, t  [0,T ], is a

Proof: See Appendix 2.A.

I

In light of this lemma, the expectation of the noise vanishes, i.e. Z

T

E

e

dt

Z

0

t

zT ðsÞsðsÞdW ðsÞdt ¼ 0; 0

after interchanging the order of integration. Thus, taking expectation in the additive utility decomposition (2.9), Z

T

e dt log CðtÞdt ¼

E 0

1 e dT

log Xð0Þ  T 1

dt þE log cðtÞ þ 1 e dðT tÞ e d 0   Q0 ðt; zðtÞ; cðtÞÞ dt: d Z

ð2:13Þ

Therefore, to maximize Z

T

e dt log CðtÞdt

E 0

over the constraint set, it suffices to maximize 4

gðt; z; cÞ ¼ log cðtÞ þ

1 1 e dðT tÞ Q0 ðt; zðtÞ; cðtÞÞ d

ð2:14Þ

26 j CHAPTER 2

pathwise over the constraint set. For a fixed path v and a time instance t, we need to solve ðP1Þ

maximize gðt; z; cÞ; pffiffiffi 4 subject to fV ðt; z; cÞ ¼ Q0 ðt; z; cÞt N 1 ðaÞkzT sðtÞk t  log

1 : 1 aV

The optimal policy for an agent maximizing her logarithmic utility of intertemporal consumption without the risk constraint is to hold the proportion fðzM ðtÞ; cM ðtÞÞgt2½0;T , 4 where cM ðtÞ ¼ d=½1 e dðT tÞ (the optimum of (P1) without the constraint is ðzM ðtÞ; cM ðtÞÞÞ. Lemma 2.2: The solution of (P1) is given by zðtÞ ¼ ð1 ^ ðbðtÞ _ 0ÞÞz ðtÞ; M

ð2:15Þ

cðtÞ ¼ uðt; ð1 ^ bðtÞÞÞcM ðtÞ1fbðtÞ > 0g ! 1 1 1fbðtÞ0g ; þ r þ log t 1 aV

ð2:16Þ

where b(t) is the root of the equation 1 fV ðt; zzM ðtÞ; uðt; zÞcM ðtÞÞ ¼ log 1 aV

ð2:17Þ

in the variable z, with 4

uðt; zÞ ¼ 1 þ

 pffiffiffi tzTM ðtÞsðtÞ N 1 ðaÞ

ð1 zÞ:

ð2:18Þ

Proof: See Appendix 2.A.

I

Theorem 2.3: To maximize the logarithmic utility of intertemporal consumption, Z

T

e dt log CðtÞ dt;

E 0

over processes ðzðtÞ; cðtÞÞ 2 FV ðtÞ, 0 5t5T, the optimal portfolio is fðzðtÞ; cðtÞÞgt2½0;T . Proof: This is a direct consequence of (2.13) and Lemma 2.2.

I

j

PORTFOLIO OPTIMIZATION UNDER THE VALUE-AT-RISK CONSTRAINT

27

Remark 1: Since at the final time cM ðT Þ ¼ 1 and cðtÞ is bounded we must have b(T) 50, so zðT Þ ¼ 0, and this means that, at the final time, the agent invests the least proportion (in absolute terms) of her wealth in stocks. By (2.15) and (2.16) it follows that zðtÞ  z ðtÞ, and cðtÞ  c ðtÞ, for any 0 5t 5T, which means that the constrained agent M M is consuming and investing less in the risky assets than the unconstrained agent. Let T1 and T2 be two final time horizons, T1
T2. Because cM ðt; T1 Þ < cM (t, T2), from Equations (2.15) and (2.16) we conclude that bðt; T1 Þ > bðt; T2 Þ, hence zðt; T1 Þ > zðt; T2 Þ, and cðt; T1 Þ > cðt; T2 Þ. Therefore, long-term agents can afford to invest more in the stock market and consume more than short-term agents (in terms of proportions).

2.5

NONLOGARITHMIC UTILITY

Let us recall that we want to maximize the expected CRRA utility (Up ðxÞ ¼ x p =p;p " 0) from intertemporal consumption and terminal wealth, Z

T

e dt Up ðCðtÞÞdt þ Ee dT Up ðX z;c ðT ÞÞ;

E

ð2:19Þ

0

over portfolio-proportion processes satisfying the Value-at-Risk constraint, i.e. ðzðtÞ; cðtÞÞ 2 FV ðtÞ, 0 5t 5T. One cannot obtain an additive decomposition into signal and noise as in the case of logarithmic utility. However, a multiplicative decomposition can be performed. By (2.7), X p ð0Þ exp Up ðX ðtÞÞ ¼ p z;c

Z

t

pQ0 ðs; zðsÞ; cðsÞÞds þ 0

Z

t T



pz ðsÞsðsÞdW ðsÞ 0

 Z t X p ð0Þ 1 2 T 2 exp pQp ðs; zðsÞ; cðsÞÞ p kz ðsÞsðsÞk ds ¼ p 2 0 ! Z t X p ð0Þ z;c N ðtÞZ z ðtÞ; pzT ðsÞsðsÞdW ðsÞ ¼ þ p 0

where 4

z;c

N ðtÞ ¼ exp

Z

t

 pQp ðs; zðsÞ; cðsÞÞds ;

ð2:20Þ

0

  Z 1 t 2 T 2 Z ðtÞ ¼ exp p jjz ðsÞsðsÞjj ds 2 0 Z t þ pzT ðsÞsðsÞdWðsÞÞ; z

4

0

with Qp defined in (2.3). By taking expectation,

ð2:21Þ

28 j CHAPTER 2

EUp ðX z;c ðtÞÞ ¼

X p ð0Þ EðN z;c ðtÞZ z ðtÞÞ: p

ð2:22Þ

The process N z,c(t) is the signal and Z z(t), a stochastic exponential, is the noise. Stochastic exponentials are local martingales, but if we impose the assumption 

 2Z T   T 2 p z ðuÞsðuÞ du < 1 E exp M 2 0

ð2:23Þ

on market coefficients (see Equation (2.11)), the process Z z(t) is a (true) martingale for all portfolio processes satisfying the constraint, as the following lemma shows. Lemma 2.4: For every ðzðtÞ; cðtÞÞ 2 FV ðtÞ the process Z z(t), t  [0, T ], is a martingale, hence EZ z ðtÞ ¼ 1. Proof: See Appendix 2.A.

I

As for utility from intertemporal consumption, Z

T

e dt Up ðCðtÞÞdt ¼

Z

T

e dt Up ðX z;c ðtÞcðtÞdt 0 Z X p ð0Þ T dt p ¼ e c ðtÞN z;c ðtÞZ z ðtÞdt: p 0

0

ð2:24Þ

We claim that Z E



T

e

dt p

z;c

z

z

c ðtÞN ðtÞðZ ðtÞ Z ðT ÞÞdt

¼ 0:

0

Indeed, by conditioning and Lemma 2.4 we obtain       E c p ðtÞN z;c ðtÞ Z z ðtÞ Z z ðT Þ ¼ E E c p ðtÞN z;c ðtÞ Z z ðtÞ Z z ðT Þ jF t    ¼ E c p ðtÞN z;c ðtÞE Z z ðtÞ Z z ðT Þ jF t ¼ 0; and Fubini’s theorem proves the claim. Hence, combined with (2.24), we obtain Z E

T

e 0

dt

Up ðCðtÞÞdt ¼

X p ð0Þ p



z

E Z ðT Þ

Z

T

e

dt p

z;c



c ðtÞN ðtÞdt :

0

The decomposition for the total expected utility (Equations (2.22) and (2.25)) is

ð2:25Þ

PORTFOLIO OPTIMIZATION UNDER THE VALUE-AT-RISK CONSTRAINT

Z

T

E

j

29

 e dt Up ðCðtÞÞdt þ Ee dT Up X z;c ðT Þ

0

¼

ð2:26Þ

X p ð0Þ  z E Z ðT ÞY z;c ðT Þ ; p

where the signal Yz, c(T) is given by Y z;c ðT Þ ¼

Z

T

e dt c p ðtÞN z;c ðtÞdt þ e dT N z;c ðT Þ;

ð2:27Þ

0

with N z, c(t) defined in (2.20). It appears natural at this point to maximize Y z, c(T) pathwise over the constraint set. For a given path v, the existence of an optimizer fðzðt; oÞ; cðt; oÞÞgt2½0;T is given by Lemma 2.5. Note that N z; c ðt; oÞ depends on the trajectory of ðzð; oÞ; cð; oÞÞ on [0,t] so one is faced with a deterministic control problem. From now on, to keep the notation simple we drop v. In the language of deterministic control we can write (2.27) as a cost functional I [x, u] given in the form I ½x; u ¼ gðxðT ÞÞ þ

Z

T

f0 ðt; xðtÞ; uðtÞÞdt;

4

gðxÞ ¼ e dT x;

ð2:28Þ

0

where u (z,c) is the control, x is the state variable, and the function 4

f0 ðt; x; uÞ ¼ e dt c p x

ð2:29Þ

A ¼ fðt; x; uÞjðt; xÞ 2 ½0; T  ð0; K ; uðtÞ 2 FV ðtÞg  Rmþ3 :

ð2:30Þ

is defined on the set

The dynamics of the state variable is given by the differential equation dx ¼ f ðt; xðtÞ; uðtÞÞ; dt

0  t  T;

ð2:31Þ

with the boundary condition x(0)1, where  2 pð p 1Þ  T   f ðt; x; uÞ ¼ x pr pc þ pz mðtÞ þ z sðtÞ : 2 4



T

ð2:32Þ

The constraints are ðt; xðtÞÞ 2 ½0; T  ð0; K and u(t)  FV(t). Due to the compactness of the set FV(t), 0 5t 5T, it follows that K B . A pair (x, u) satisfying the above conditions is called admissible. The problem of finding the maxima of I [x, u] within all admissible pairs (x, u) is called the Bolza control problem. Classical existence theory for deterministic control does not apply to the present situation and we proceed with a direct proof of existence.

30 j CHAPTER 2 4 Lemma 2.5: There exists a solution f uðtÞg0tT ¼ fðzðtÞ; cðtÞÞg0tT for the Bolza control problem defined above.

Proof: See Appendix 2.A.

I

An optimal solution f uðtÞg0tT ¼ fðzðtÞ; cðtÞÞg0tT is characterized by a system of forward backward equations (also known as the Pontryagin maximum principle). Let ~ l ¼ ðl0 ; l1 Þ be the adjoint variable and 4

Hðt; x; u; ~ lÞ ¼ l0 f0 ðt; x; uÞ þ l1 f ðt; x; uÞ the Hamiltonian function. The necessary conditions for the Bolza control problem (the Pontryagin maximum principle) can be found in Cesari (1983) (Theorem 2.5.1.i). In general, they are not sufficient for optimality. We prove that, in our context, the necessary conditions are also sufficient, as the following lemma shows. Lemma 2.6: A pair xðtÞ; uðtÞ ¼ ðzðtÞ; cðtÞÞ 2 FV ðtÞ, 05t 5T, is optimal, i.e. it gives the maximum for the functional I [x, u], if and only if there is an absolutely continuous non-zero vector function of Lagrange multipliers  l ¼ ðl0 ; l1 Þ, 05t5T, with l0 a constant, l0 ]0, such 4 that the function MðtÞ ¼ Hðt; xðtÞ; uðtÞ; lðtÞÞ is absolutely continuous and one has 1. adjoint equations: dM ¼ Ht ðt; xðtÞ; uðtÞ; lðtÞÞ a:e:; dt dl1 dt

¼ Hx ðt; xðtÞ; uðtÞ; lðtÞÞ a:e:;

ð2:33Þ

ð2:34Þ

2. maximum condition: uðtÞ 2 arg maxv2FV ðtÞ Hðt; xðtÞ; v; lðtÞÞ a:e:;

ð2:35Þ

l1 ðT Þ ¼ l0 g 0 ð x ðT ÞÞ:

ð2:36Þ

3. transversality:

Proof: See Appendix 2.A.

I

The following technical requirement on the market coefficients is sufficient to make  fðzðtÞ; cðtÞÞgt2½0;T an optimal portfolio process for maximizing the CRRA utility under the Value-at-Risk constraint, as Theorem 2.7 shows. We assume that market coefficients are

PORTFOLIO OPTIMIZATION UNDER THE VALUE-AT-RISK CONSTRAINT

j

31

totally unhedgeble, i.e. the mean rate of return process faðtÞgt2½0;T and the matrix-valued  tg volatility process fsðtÞgt2½0;T are adapted to filtration fF 0t1 generated by a Brownian motion independent of the Brownian motion driving the stocks (see Equation (2.1)). Theorem 2.7: A solution for maximizing Z

T

e dt Up ðCðtÞÞ dt þ E e dT Up ðXðT ÞÞ

E 0

over processes ðzðtÞ; cðtÞÞ 2 FV ðtÞ, 0 5t 5T, is a process fðzðtÞ; cðtÞÞgt2½0;T , which, on every v, solves (2.33), (2.34), (2.35) and (2.36). Proof: Lemma 2.5 gives the existence, on every v, of fðzðtÞ; cðtÞÞgt2½0;T , optimal for the Bolza control problem, i.e. it maximizes Yz,c(T) defined in (2.27) over FV(t), 0 5t 5T. According to Lemma 2.5 it should solve (2.33), (2.34), (2.35) and (2.36) pathwise and these equations are sufficient for optimality. Let ðzðtÞ; cðtÞÞ 2 FV ðtÞ be another control. Let   Zz(t), Z z ðtÞ and Yz,c(T), Y z; cðT Þ be as in (2.21) and (2.27). The processes fZ z ðtÞg0tT ,   tg fZ z ðtÞg0tT are martingales by Lemma 2.4. Moreover, the independence of fF 0t1 and fF t g0t1 implies h i      ðT Þ ¼ E Z z ðT ÞF  ðT Þ ¼ 1: E Z z ðT ÞF   ðT Þ. Therefore, by (2.26) Lemma 2.6 shows that Y z;c ðT Þ is measurable with respect to F and iterated conditioning

Z

T

E

   e dt Up ðCðtÞÞdt þ Ee dT Up X z;c ðT Þ

0

¼

X p ð0Þ  z  E Z ðT ÞY z;c ðT Þ p

¼

 X p ð0Þ   z   ðT Þ E E Z ðT ÞY z;c ðT Þ j F p

¼

   X p ð0Þ  z;c  ðT Þ E Y ðT ÞE Z z ðT Þ j F p

¼

X p ð0Þ z;c EY ðT Þ: p

Since ðzðtÞ; cðtÞÞ maximizes Y z;c ðT Þ over the constraint set,

32 j CHAPTER 2

Z

T

e dt Up ðCðtÞÞdt þ Ee dT Up ðX z;c ðT ÞÞ

E 0

¼

X p ð0Þ EðZ z ðT ÞY z;c ðT ÞÞ p



X p ð0Þ  EðZ z ðT ÞY z;c ðT ÞÞ p

¼

X p ð0Þ   ðT Þ Þ EðE½Z z ðT ÞY z;c ðT Þ j F p

¼

X p ð0Þ   ðT Þ Þ EðY z;c ðT ÞE½Z z ðT Þ j F p

¼

X p ð0Þ z;c EY ðT Þ: p

Remark 1: Let the interest rate and the discount factor be stochastic processes. In the formulae of fV(t, z, c) and Qp(t, z, c) the constant r is replaced by r(t). All the results remain true if we assume that frðtÞg0t1 and fdðtÞg0t1 are non-negative uniformly bounded  continuous processes adapted to FðtÞ. We have considered the case of Value-at-Risk (VaR) in defining the risk constraint. The same methodology applies if one considers other measures of risk, as long as the corresponding constraint set is convex and compact.

2.6

NUMERICAL SOLUTION

Theorem 2.7 shows that the solution for every path of a Bolza control problem yields the optimal portfolio proportion for a VaR-constrained agent. The solution exists and is characterized by a system of forward backward equations that are also sufficient for optimality. In this section, by an appropriate discretization of control and state variables, the Bolza control problem is transformed into a finite-dimensional nonlinear program that can be solved by standard sequential quadratic programming (SQP) methods. The first step is to transform the Bolza problem into a Mayer control problem by introducing a new state variable x0, with the boundary condition x0(0)  0 and an additional differential equation, dx 0 dt

¼ f0 ðt; xðtÞ; uðtÞÞ:

The cost functional is then I ½x; u ¼ x 0 ðT Þ þ gðxðT ÞÞ (see Equation (2.28)). Let us denote as y ¼ ðx 0 ; xÞ the vector of state variables that satisfy the differential equation dy dt

¼ f~ðt; xðtÞ; uðtÞÞ;

PORTFOLIO OPTIMIZATION UNDER THE VALUE-AT-RISK CONSTRAINT

j

33

with f~ ¼ ð f0 ; f Þ (see Equations (2.29) and (2.31)). The following discretization scheme is taken from Stryk (1993). The novel feature here is that the necessary first-order conditions of the discretized problem converge to the necessary first-order conditions of the continuous problem. A partition of the time interval 0 ¼ t1 < t2 <    < tN ¼ T is chosen. The parameters Y of the nonlinear program are the values of the control and state variables at the grod points tj, j ¼ 1; . . . ; N , and the final time tN  T, Y ¼ ðuðt1 Þ; . . . ; uðtN Þ; yðt1 Þ; . . . ; yðtN Þ; tN Þ 2 R4Nþ1 : The controls are chosen as piecewise linear interpolating functions between u(tj) and uðtjþ1 Þ, for tj  t < tjþ1, uapp ðtÞ ¼ uðtj Þ þ

t tj tjþ1 tj

ðuðtjþ1 Þ uðtj ÞÞ:

The states are chosen as continuously differentiable functions and piecewise cubic Hermite polynomials between y(tj) and yðtjþ1 Þ, with y_app ðsÞ ¼ f~ðxðsÞ; uðsÞ; sÞ at s ¼ tj ; tjþ1 , yapp ðtÞ ¼

3 X

j dk

t tj

k¼0

hj

!k ; ð2:37Þ

tj  t < tjþ1 ; j ¼ 1; . . . ; N 1; j

d0 ¼ yðtj Þ;

j d1 ¼ hj f~j ;

ð2:38Þ

j d2 ¼ 3yðtj Þ 2hj f~j þ 3yðtjþ1 Þ hj f~jþ1 ;

ð2:39Þ

j d3 ¼ 2yðtjþ1 Þ þ hj f~j 2yðtjþ1 Þ þ hj f~jþ1 ;

ð2:40Þ

where 4 f~j ¼ f~ðxðtj Þ; uðtj Þ; tj Þ;

4

hj ¼ tjþ1 tj :

The reader can learn more about this in common textbooks of Numerical Analysis such as, Stoer and Bulrisch (1983). This way of discretizing has two advantages. The number of _ j Þ; j ¼ 1; . . . ; N , are not parameters of the nonlinear program is reduced because yðt

34 j CHAPTER 2

parameters and the number of constraints is reduced because the constraints y_app ðtj Þ ¼ f~ðxðtj Þ; uðtj Þ; tj Þ, j ¼ 1; . . . ; N , are already fulfilled. We impose the Value-at-Risk constraint (see Equation (2.6)) at the grid points fV ðtj ; uðtj ÞÞ  log

1 ; 1 aV

u ¼ ðz; cÞ; j ¼ 1; . . . ; N :

ð2:41Þ

Another constraint imposed is the so-called collocation constraint f~ðxðt~j Þ; uðt~j Þ; t~j Þ ¼ y_app ðt~j Þ;

j ¼ 1; . . . ; N ;

0 ðt~j Þ; f0 ðxðt~j Þ; uðt~j Þ; t~j Þ ¼ x_app

j ¼ 1; . . . ; N ;

ð2:42Þ

f ðxðt~j Þ; uðt~j Þ; t~j Þ ¼ x_app ðt~j Þ;

j ¼ 1; . . . ; N ;

ð2:43Þ

or componentwise

and

4

where t~j ¼ ðtj þ tjþ1 Þ=2, and the boundary condition y(0)  (0,1). The nonlinear program is to maximize I [yN , tN] subject to constraints (2.41), (2.42) and (2.43). It can be solved using NPSOL, a set of Fortran subroutines developed by Gill et al. (1986). NPSOL uses a sequential quadratic programming (SQP) algorithm, in which the search direction is the solution of a quadratic programming (QP) subproblem. The Lagrangian of the nonlinear program is LðY ; f; uÞ ¼ I ½yN ; tN þ

N X j¼1

þ

N 1 X

uj

1 fV ðtj ; uðtj ÞÞ log 1 aV

!

0 f0j ðf0 ðxðt~j Þ; uðt~j Þ; t~j Þ x_app ðt~j ÞÞ

j¼1

þ

N 1 X

f1j ðf ðxðt~j Þ; uðt~j Þ; t~j Þ x_app ðt~j ÞÞ;

j¼1

where u ¼ ðu1 ; . . . ; uN Þ 2 R N , f0 ¼ ðf01 ; . . . ; f0N 1 Þ 2 R N 1 and f1 ¼ ðf11 ; . . . ; f1N 1 Þ 2 4 4 4 R N 1 are the shadow prices. Let us denote zðti Þ ¼ zi , cðti Þ ¼ ci and xðti Þ ¼ xi . A solution of the nonlinear program satisfies the necessary first-order conditions of Karush, Kuhn and Tucker @L @zi

¼ 0;

@L @ci

¼ 0;

@L @xi

¼ 0;

ð2:44Þ

i ¼ 1; . . . ; N : The necessary first-order optimality conditions of the continuous problem are obtained in 4 the limit from (2.44) as follows. Let h ¼ maxfhj ¼ tjþ1 tj : j ¼ 1; . . . ; N 1g be the norm

PORTFOLIO OPTIMIZATION UNDER THE VALUE-AT-RISK CONSTRAINT

j

35

of the partition. Letting h 00, after some calculations (see p. 5 of Stryk 1993) it is shown that, at t  ti, @L 3 @f ðxðti Þ; uðti Þ; ti Þ 3 0 @f0 ðxðti Þ; uðti Þ; ti Þ þ fi

! f1i @zi 2 @z 2 @z @f ðuðti Þ; ti Þ ; þ ui V @z @L 3 @f ðxðti Þ; uðti Þ; ti Þ 3 0 @f0 ðxðti Þ; uðti Þ; ti Þ þ fi

! f1i @ci 2 @c 2 @c @f ðuðti Þ; ti Þ ; þ ui V @c @L 3 3 @f ðxðti Þ; uðti Þ; ti Þ 3 0 @f0 ðxðti Þ; uðti Þ; ti Þ þ fi :

! f_1i þ f1i @xi 2 2 @x 2 @x Therefore, @L=@zi ¼ 0 and @L=@ci ¼ 0 converge to an equation equivalent to the maximum condition (2.35), and @L=@xi ¼ 0 converges to the adjoint Equation (2.34). This discretization scheme gives good estimates for the adjoint variables. In the following we perform some numerical experiments. We consider one stock following a geometric Brownian motion with drift a1  0.12 and volatility s  0.2. The choices for the horizon t and the confidence level a are largely arbitrary, although the Basle Committee proposals of April 1995 prescribed that VaR computations for the purpose of assessing bank capital requirements should be based on a uniform horizon of 10 trading days (two calendar weeks) and a 99) confidence level (Jorion 1997). We take t  1u25, a  0.01, the interest rate r  0.05 and the discount factor d  0.1 (Figure 2.1).

2.7

CONCLUDING REMARKS

Let us summarize the results. This chapter examines in a stochastic paradigm the portfolio choice problem under a risk constraint, which is applied dynamically consistent at every time instant. The classical stochastic control methods, Dynamic Programming and the Martingale Method, are not very effective in this context. The latter works if the risk constraint is imposed in a static way. The Dynamic Programming approach (as shown in Section 2.3) leads to a highly nonlinear PDE. If the agent has CRRA preferences we propose a new method that relies on a decomposition of the utility into signal and noise. We neglect the noise (the expectation operator takes care of this) and this leads to a deterministic control problem on every path. We have reported explicit analytical solutions for the case of logarithmic utility even if the market coefficients are random processes. In this case, on every path the deterministic control problem is just a timedependent constrained nonlinear program. The explicit solution shows that constrained agents consume and invest less in stocks than unconstrained agents, and long-term agents invest and consume more than short-term agents. These effects support the use of dynamically consistent risk constraints. If the utility is non-logarithmic CRRA we have to analyse a Bolza control problem on every path. We still allow the market coefficients to be

36 j CHAPTER 2 0.9 Unconstrained Constrained

0.8

Unconstrained Constrained

1

0.7

0.8

0.6 0.5

0.6

0.4 0.4

0.3 0.2

0.2

0.1 0

0

0.5

1

1.5 Graph 1

2

2.5

3

0

0.5

1

1.5 Graph 2

2

2.5

3

4

Unconstrained Constrained

1.2

0

3.5 3

1

Unconstrained Constrained

2.5

0.8

2 0.6 1.5 0.4

1

0.2

0.5

0

0

0.5

1

1.5 Graph 3

2

2.5

3

0

0

0.5

1

1.5 Graph 4

2

2.5

3

FIGURE 2.1 Asset allocation with and without VaR constraints, for the utility maximization of intertemporal consumption and final wealth. The graphs corresponds to different values of CRRA, p : p  1.5 (Graph 1), p  1 (Graph 2), p 0.5 (Graph 3) and p  0.5 (Graph 4). The x axis represents the time and the y axis the proportion of wealth invested in stocks. Note that the Merton line and, as time goes by, the portfolio value, increase, hence the VaR constraint becomes binding and reduces the investment in the risky asset. At the final time the agent is investing the least in stocks (in terms of proportions). When p increases, i.e. when the agent becomes less risk-averse, the effect of the VaR constraint becomes more significant.

random, but independent of the Brownian motion driving the stocks. Theorem 2.7 shows that a solution of the deterministic control problem is an optimal policy. Although the existence of an optimal policy is known if the constraint set is convex (Cvitanic and Karatzas 1992), it does not necessarily yield the existence for the Bolza problem. Standard existence theorems do not apply, but we manage to give a direct proof of existence in Lemma 2.5. The solution of the Bolza problem must solve a system of forward backward equations (the first-order necessary conditions) and this is also sufficient for optimality. In Section 2.6 we suggest a numerical treatment of the Bolza problem. The reduction of the stochastic control problem to a deterministic problem relies on the structure of the CRRA preferences. It would be interesting to extend this to other classes of preferences, because it turns out to be very effective for the case of dynamically consistent risk constraints.

ACKNOWLEDGEMENTS The author would like to thank Ulrich Haussmann, Philip Loewen, Karsten Theissen, and two anonymous referees for helpful discussions and comments. This work was supported by NSERC under research grant 88051 and NCE grant 30354 (MITACS).

PORTFOLIO OPTIMIZATION UNDER THE VALUE-AT-RISK CONSTRAINT

j

37

REFERENCES Atkinson, C. and Papakokinou, M., Theory of optimal consumption and portfolio selection under a Capital-at-Risk (CaR) and a Value-at-Risk (VaR) constraint. IMA J. Mgmt. Math., 2005, 16, 37/70. Basak, S. and Shapiro, A., Value-at-risk-based risk management: Optimal policies and asset prices. Rev. Finan. Stud., 2001, 14, 371 /405. Berkelaar, A., Cumperayot, P. and Kouwenberg, R., The effect of VaR-based risk management on asset prices and volatility smile. Eur. Finan. Mgmt., 2002, 8, 139/164; 1, 65 /78. Cesari, L., Optimization—Theory and Applications Problems with Ordinary Differential Equations, 1983 (Springer: New York). Cuoco, D. and Liu, H., An analysis of VaR-based capital requirements. Preprint, Finance Department, The Wharton School, University of Pennsylvania, 2003. Cuoco, D., Hua, H. and Issaenko, S., Optimal dynamic trading strategies with risk limits. Preprint, Yale International Center for Finance, 2001. Cvitanic, J. and Karatzas, I., Convex duality in convex portfolio optimization. Ann. Appl. Probabil., 1992, 3, 652/681. Delbaen, F. and Schachermayer, W., A general version of the fundamental theorem of asset pricing. Math. Ann., 1994, 300, 463/520. Emmer, S., Klu¨ ppelberg, C. and Korn, R., Optimal portfolios with bounded capital at risk. Math. Finan., 2001, 11, 365/384. Gill, P.E., Murray, W., Saunders, M.A. and Wright, M.H., User’s guide for NPSOL (version 4.0). Report SOL 86-2, Department of Operations Research, Stanford University, 1986. Jorion, P., Value at Risk, 1997 (Irwin: Chicago). Karatzas, I. and Shreve, S.E., Brownian Motion and Stochastic Calculus, 2nd ed., 1991 (Springer: New York). Karatzas, I. and Shreve, S.E., Methods of Mathematical Finance, 1998 (Springer: New York). Leippold, M., Trojani, F. and Vanini, P., Equilibrium impact of value-at-risk. Preprint, Swiss Banking Institute, University of Zurich, Switzerland, 2002. Pirvu, T.A., Maximizing portfolio growth rate under risk constraints. Doctoral Dissertation, Department of Mathematics, Carnegie Mellon University, 2005. Stoer, J. and Bulrisch, R., Introduction to Numerical Analysis, 2nd ed., 1983 (Springer: New York). Stryk, O., Numerical solution of optimal control problems by direct collocation. In Optimal Control—Calculus of Variations, Optimal Control Theory and Numerical Methods. International Series of Numerical Mathematics, Vol. 111, 1993, pp. 129–143 (Birkhauser: Basle and Boston). Vidovic, D.G., Lavassani, L.A., Li, X. and Ware, A., Dynamic portfolio selection under capital-atrisk. University of Calgary Yellow Series, Report 833, 2003. Yiu, K.F.C., Optimal portfolios under a Value-at-Risk constraint. J. Econ. Dynam. Contr., 2004, 28, 1317/1334.

APPENDIX 2.A Proof R t T (proof of Lemma 2.1): In order to prove the martingale property of z ðsÞsðsÞdW ðsÞ it suffices to show that 0 Z

T

E 0

kzT ðuÞsðuÞk2 du < 1:

ðA2:1Þ

38 j CHAPTER 2

Note that  T    z ðtÞmðtÞ ¼ zT ðtÞsðtÞsT ðtÞz ðtÞ M      zT ðtÞsðtÞ  zT ðtÞsðtÞ:

ðA2:2Þ

M

For ðzðtÞ; cðtÞÞ 2 FV ðtÞ, we have 

 2 1 T   t r cðtÞ þ z ðtÞmðtÞ z ðtÞsðtÞ 2 pffiffiffi þ N 1 ðaÞkzT ðtÞsðtÞk t  logð1 aV Þ: T

This combined with (A2.2) yields  T    z ðtÞsðtÞ  k _ zT ðtÞsðtÞ; 1 M

ðA2:3Þ

for some positive constant k1, where, as usual, a _ b ¼ maxða; bÞ. In light of assumption (2.12) the inequality (A2.1) follows. I Proof (proof of Lemma 2.2): The proof relies on the method of Lagrange multipliers. The concave function 1 4 gðt; z; cÞ ¼ log c þ ð1 e dðT tÞ ÞQ0 ðt; z; cÞ d

ðA2:4Þ

is maximized over ðz; cÞ 2 Rm  ½0; 1Þ by ðzM ðtÞ; cM ðtÞÞ, where 4 4

1 zM ðtÞ ¼ ðsðtÞsT ðtÞÞ mðtÞ and cM ðtÞ ¼ d=½1 e dðT tÞ . If this point is in the Value-atRisk constraint set, then is the optimal solution of (P 1). Otherwise, the concave function g is maximized over the compact, convex set FV(t) at a unique point ðzðtÞ; cðtÞ); moreover, this point must be on the boundary of FV(t). Hence, it solves the optimization problem (P2)

maximize gðt; z; cÞ; pffiffiffi 4 subject to fV ðt; z; cÞ ¼ Q0 ðt; z; cÞt N 1 ðaÞkzT sðtÞk t ¼ log

1 : 1 aV

The function fV is not differentiable when z is the zero vector. Let us assume that the optimal zðtÞ is not the zero vector. According to the Lagrange multiplier theorem, either HfV ðt; zðtÞ; cðtÞÞ is the zero vector or there is a positive l such that Hgðt; zðtÞ; cðtÞÞ ¼ lHfV ðt; zðtÞ; cðtÞÞ:

ðA2:5Þ

The first case cannot occur and computations show that zðtÞ is parallel to zM(t). This implies that the optimal solution ðzðtÞ; cðtÞÞ ¼ ðl1 zM ðtÞ; l2 cM ðtÞÞ, with l1 ; l2 the solution of

PORTFOLIO OPTIMIZATION UNDER THE VALUE-AT-RISK CONSTRAINT

(P3)

j

39

maximize lðl1 ; l2 Þ; subject to fV ðt; l1 zM ðtÞ; l2 cM ðtÞÞ ¼ log

1 ; 1 aV

where 4

lðl1 ; l2 Þ ¼ gðt; l1 zM ðtÞ; l2 cM ðtÞÞ:

ðA2:6Þ

The concave function l is maximized over R2 at (1,1). We know that this point is not in the = FV ðtÞ), hence l1 < 1, l2 < 1 constraint set (this is because we have assumed ðzM ðtÞ; cM ðtÞÞ 2     and either HfV ðt; l1 zM ðtÞ; l2 cM ðtÞÞ ¼ 0 or Hlðl1 ; l2 Þ ¼ gHfV ðt; l1 zM ðtÞ; l2 cM ðtÞÞ, for some positive Lagrange multiplier g. The first case cannot occur and by eliminating g we obtain  l2 ¼ uðt;  l1 Þ, where u was defined in (2.18). Henceforth, l1 is the unique root of the equation fV ðt; zzM ðtÞ; uðt; zÞcM ðtÞÞ ¼ log

1 1 aV

ðA2:7Þ

in the variable z. It may be the case that the root of this equation is negative, in which case ! 1 1 ; 0m ; r þ log t 1 aV

ðzðtÞ; cðtÞÞ ¼

where 0m is the m-dimensional vector of zeros. Proof (proof of Lemma 2.4): Assumption (2.23) combined with (A2.3) and the Novikov condition (see Karatzas and Shreve 1991, p. 199, Corollary 5.13) make the process Z z(t) a martingale. I Proof (proof of Lemma 2.5): According to (2.31) and (2.32), xðtÞ ¼ exp

Z



t

f1 ðt; cðuÞ; zðuÞÞdu ;

ðA2:8Þ

0

with 4

f1 ðt; c; zÞ ¼ pr pc þ pzT mðtÞ þ

 pð p 1Þ  zT sðtÞ2 : 2

Let us recall that, for uðtÞ ¼ ðzðtÞ; cðtÞÞ 2 FV ðtÞ,     T z ðtÞsðtÞ  k _ zT ðtÞsðtÞ; 1 M

ðA2:9Þ

40 j CHAPTER 2

hence kzT ðtÞsðtÞk is uniformly bounded on [0, T ] due to the continuity of market coefficients (see Equation (2.11)). Moreover, one can conclude that kzðtÞk  K1 and _  K2 on cðtÞ  K1 for some constant K1. Gronwall’s lemma gives xðtÞ  K2 and jxðtÞj _ [0, T] (here, xðtÞ ¼ dx=dtÞ. Let (x_n, u_n) be a maximizing sequence, i.e. I ½xn ; un

! sup I ½x; u . The above arguments show that the sequence xn is uniformly bounded and equicontinuous, thus by the Arzela Ascoli theorem it converges uniformly to some function x~. According to Komlos’ Lemma (see Lemma A1.1 of Delbaen and Schachermayer 1994) we can find some sequences of convex combinations z 2 convðz ; z ; Þ and c 2 convðc ; c ; Þ that converge a.e. to some measurable n n nþ1 n n nþ1 4 functions z and c. Moreover, uðtÞ ¼ ðzðtÞ; cðtÞÞ 2 FV ðtÞ, 0 5t 5T, due to the convexity and compactness of the set FV(t). Let us denote xn as the sequence of state variables corresponding to these controls, i.e. /

xn ðtÞ ¼ exp

Z

t

 f1 ðt; cn ðsÞ; zðsÞÞds ;

0t T

0

(see Equation (A2.8)). Let us assume p
0; the case pB0 can be treated similarly. Due to the concavity of the function f1, ln xn ðtÞ  convðln xn ðtÞ; ln xnþ1 ðtÞ; . . .Þ, where the convex 4 combination is that defining z , c . If y ¼ expðconvðln x ðtÞ; ln x ðtÞ; . . .ÞÞ, then x ðtÞ  n

n

n

n

nþ1

n

yn ðtÞ and yn ðtÞ !~ x ðtÞ, i.e. yn ðtÞ xn ðtÞ ! 0 for t  [0,T ]. By the dominated convergence x ðtÞ, 0 5t 5T, a.e., where theorem xn ðtÞ !

xðtÞ ¼ exp

Z

t

  f1 ðt; cðsÞ; zðsÞÞds ;

0  t  T:

0

From Fatou’s lemma, the dominated convergence theorem and the concavity of the function f0 in u (see Equation (2.29)), it follows that yn ; un

I ½ x ; u  lim sup I ½ xn ; un  lim sup I ½ ¼ lim sup I ½xn ; un ¼ sup I ½x; u :

Proof (proof of Lemma 2.6): By Theorem 5.1.i of Cesari (1983), fðzðtÞ; cðtÞÞgt2½0;T

should solve (2.33), (2.34), (2.35) and (2.36). For sufficiency, let us consider l0 ¼ 1, and define the maximized Hamiltonian 4 lÞ ¼ max Hðt; x; v; lÞ: H  ðt; x;  v2FV ðtÞ

Let (x(t), u(t)) be another admissible pair. Since the Hamiltonian is linear in x, by the adjoint equation for l1 and the maximum condition

PORTFOLIO OPTIMIZATION UNDER THE VALUE-AT-RISK CONSTRAINT

Hðt; xðtÞ; uðtÞ;  lðtÞÞ Hðt; xðtÞ; uðtÞ; lðtÞÞ  H  ðt; xðtÞ; lðtÞÞ H  ðt; x; lðtÞÞ ¼

l01 ðtÞðxðtÞ

Z

T

j

41

ðA2:10Þ

xðtÞÞ;

one has I ½ x ; u I ½x; u ¼

0

þ

ðHðt; xðtÞ; uðtÞ; lðtÞÞ Hðt; xðtÞ; uðtÞ; lðtÞÞÞdt

Z

T

_ x_ ðtÞÞdt gðxðT ÞÞ þ gð l1 ðtÞðxðtÞ x ðT ÞÞ: 0

The inequality (A2.10) and the transversality condition (2.36) yield I ½x; u I ½x; u 

Z

T

l01 ðtÞðxðtÞ xðtÞÞdt 0

þ

Z

T

_ x_ ðtÞÞdt gðxðT ÞÞ þ gð l1 ðtÞðxðtÞ x ðT ÞÞ 0

¼ l1 ðT ÞðxðT Þ xðT ÞÞ gðxðT ÞÞ þ gð x ðT ÞÞ x ðT ÞÞðxðT Þ xðT ÞÞ gðxðT ÞÞ þ gð x ðT ÞÞ ¼ g 0 ð ¼ gðxðT ÞÞ gð x ðT ÞÞ gðxðT ÞÞ þ gð x ðT ÞÞ ¼ 0; proving the optimality of ð x ðtÞ; uðtÞÞ.

I

CHAPTER

3

Dynamic Consumption and Asset Allocation with Derivative Securities YUAN-HUNG HSUKU CONTENTS 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 3.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 3.2.1 Investment Opportunity Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .46 3.2.2 Preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .48 3.3 Optimal Consumption Policy and Dynamic Asset Allocation Strategies . . . . 49 3.3.1 Optimal Consumption and Dynamic Asset Allocation Strategies . . . .49 3.3.2 Optimal Consumption and Dynamic Asset Allocation Strategies with Time-Varying Risk Premiums . . . . . . . . . . . . . . . . . .52 3.4 Analyses of the Optimal Consumption Rule and Dynamic Asset Allocation Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 Appendix 3.A: Derivation of the Exact Solution When j  1. . . . . . . . . . . . . . 62 Appendix 3.B: Derivation of the Approximate Results . . . . . . . . . . . . . . . . . . . 64

3.1

T

INTRODUCTION

consumption and dynamic asset allocation of stocks, bonds and derivatives for long-term investors in contrast to the standard optimal dynamic asset allocation strategies involving only stocks and bonds. The chapter explores and attempts to understand the effect of introducing a non-redundant derivative security on an already-existing stock—in particular, on the volatility of stock returns. Instead of a single-period (two-date) result, we also delve into the optimal intertemporal consumption as well as dynamic asset allocation strategies under a stochastic investment opportunity set. HIS PAPER FINDS THE OPTIMAL

43

44 j CHAPTER 3

We show that these long-term investors have access to an environment where investment opportunities vary over time with stochastic volatility. There is abundant empirical evidence that the conditional variance of stock returns is not constant over time. Merton (1969, 1971, 1973a) shows that if investment opportunities vary over time, then multi-period investors will have a very different optimal consumption rule and portfolio strategies than those of single-period investors. If multi-period investors hope to maintain a stable long-run consumption stream, then they may seek to hedge their exposures to future shifts in the time-varying investment opportunity set, and this creates extra intertemporal hedging demands for financial assets. Following Merton’s (1969, 1971) introduction of the standard intertemporal consumption and investment model, it has been studied extensively in the finance literature and has become a classical problem in financial economics. The literature on the broad set of issues covering intertemporal consumption and investment or optimal dynamic asset allocation strategies restricts investors’ access to bond and stock markets only, while excluding the derivatives market. Haugh and Lo (2001), Liu and Pan (2003) and Ilhan et al. (2005) remove this restriction and introduce derivative securities in the financial market. Cont et al. (2005) consider the problem of hedging a contingent claim in a market where prices of traded assets can undergo jumps, by trading in the underlying asset and a set of traded options, and they work in a continuous time setting. However, they give an expression for the hedging strategy which minimizes the variance of the hedging error, while our model provides an investor who aims to maximize expected utility and gives an expression for the dynamic asset allocation strategies. They take a look at optimal dynamic or buy-and-hold portfolio strategies for an investor who can control not just the holding of stocks and bonds, but also derivatives. Our work is related to their research and makes several extensions. The classical option pricing framework of Black and Scholes (1973) and Merton (1973) is based on a replication argument in a continuous trading model with no transaction costs. The presence of transaction costs, however, invalidates the Black
Scholes arbitrage argument for option pricing, since continuous revision requires that the portfolio be rebalanced infinitely, implying infinite transaction costs (Chalasani and Jha 2001). As a result, some of the recent literature has begun to work on transactions costs. However, this is beyond the scope of this chapter which follows Black and Scholes (1973) and Merton (1973) in assuming zero transaction costs for stocks and options trading. First, Cont et al. (2005) solve the optimal investment or asset allocation problem of a representative investor whose investment opportunity includes not only the usual riskless bond and risky stock, but also derivatives on the stock. In their model, investors are assumed to have a specified utility defined over wealth at a single terminal date. This chapter extends this setting to consider a model in which a long-term investor chooses consumption as well as an optimal portfolio including the riskless bond, risky stock and derivatives on the stock. We then maximize a utility function defined over intermediate consumption rather than terminal wealth. Abstraction from the choice of consumption over time implies that investors value only wealth at a single terminal date, no consumption takes place before the terminal date, and all portfolio returns are reinvested until that date. The assumption that investors derive /

DYNAMIC CONSUMPTION AND ASSET ALLOCATION

j

45

utility only from terminal wealth and not from intermediate consumption simplifies the analysis through avoidance of an additional source of non-linearity in the differential equation. However, many long-term investors seek a stable consumption path over a long horizon. This simplification makes it hard to apply the above-mentioned papers’ results to the realistic problem facing an investor saving for the future. Very often, intermediate consumption can be used as an indicator of marginal utility, especially in the asset pricing related literature (Campbell and Viceira 1999). The second extension of this chapter is in contrast to Liu and Pan (2003), in which investors are assumed to have power utility, in contrast to Ilhan et al. (2005), in which investors maximize expected exponential utility of terminal wealth and also in contrast to Haugh and Lo (2001), who set the special cases of CRRA and CARA preferences. In this model we assume that investors have continuous-time recursive preferences introduced by Duffie and Epstein (1992b). This allows us not only to provide the effects of risk aversion, but also to separate cleanly an investor’s elasticity of intertemporal substitution in consumption from the coefficient of relative risk aversion. This is because power utility functions restrict risk aversion to be the reciprocal of the elasticity of intertemporal substitution, but in fact these parameters have very different effects on optimal consumption and portfolio choice (Campbell and Viceira 1999; Bhamra and Uppal 2003; Chacko and Viceira 2005). Under the model settings of Liu and Pan (2003), the mean
variance allocation to stocks, i.e. the ratio between expected stock excess returns and stock return volatility, is constant, while our work more realistically reflects the real-world situation with timevarying allocation components. Liu and Pan (2003) consider the Heston (1993) and Stein and Stein (1991) models of stochastic volatility, in which volatility follows a meanreverting process and stock returns are a linear function of volatility. Our setting is the more general assumption that expected stock returns are an affine function of volatility. In this setting, the Liu and Pan (2003) result is the special case where we constrain the intercept of the affine function to be zero. Therefore, we provide a dynamic asset allocation in both stocks and derivatives, in contrast both to Ilhan et al. (2005), who restrict to a static position in derivative securities, and to Haugh and Lo (2001), who construct a buy-and-hold portfolio of stocks, bonds and options that involves no trading once established at the start of the investment horizon. A buy-and-hold portfolio with a derivative securities strategy may come closest to optimal dynamic asset-allocation policies involving only stocks and bonds, as concluded by Haugh and Lo (2001). We know that the problem of dynamic investment policies, i.e. asset-allocation rules, arise from standard dynamic optimization problems in which an investor maximizes the expected utility as shown by Haugh and Lo (2001). At the same time they pose the following problem: given an investor’s optimal dynamic investment policy for two assets, stocks and bonds, construct a ‘buy-and-hold’ portfolio—a portfolio that involves no trading once it is established—of stocks, bonds and options at the start of the investment horizon. They state that this comes closest to the optimal dynamic policy (Haugh and Lo 2001), but their strategies differ from our dynamic asset allocation strategies which involve not only stocks, but also derivative securities. /

46 j CHAPTER 3

Chacko and Viceira (2005) examine the optimal consumption and portfolio choice problem of long-horizon investors who only have access to a riskless asset with constant return and a risky asset (stocks), without introducing any derivative securities. Comparing with Chacko and Viceira (2005), our generalized model considers that holding derivative securities complicates the asset allocation strategies for long-horizon investors. If one does not hold any derivative securities as in Chacko and Viceira (2005), the assumption of imperfect instantaneous correlation between risky stock returns and its stochastic volatility means that the intertemporal hedging component of the risky stock can only provide partial hedging ability for multi-period investors when facing the time-varying investment opportunity set. In this chapter, where we introduce non-redundant derivative securities written on the risky stock in the incomplete financial market under this optimal dynamic asset allocation, the derivative securities in the asset allocation can provide differential exposures to stochastic volatility and make the market complete. The derivative securities can also supplement the deficient hedging ability of the intertemporal hedging component of the risky stock, because of the nonlinear nature of derivative securities. The chapter will obtain a solution to this problem which is exact for investors with unit elasticity of intertemporal substitution of consumption, and approximate otherwise. The chapter is organized as follows. Section 3.2 describes the model and environment assumed in this chapter. Section 3.3 develops the model of optimal consumption policy and dynamic asset allocation strategies. We also extend the model with constant expected excess returns and constant volatility risk premiums to time-varying instantaneous expected excess returns in relation to the risky stock and time-varying stochastic volatility risk premium. Section 3.4 provides the analyses of optimal consumption policy and dynamic asset allocation strategies. Finally, conclusions are given in Section 3.5.

3.2

THE MODEL

3.2.1

Investment Opportunity Set

In this chapter we assume that wealth comprises investments in traded assets only. There are two prime assets available for trading in the economy. One of the assets is an instantaneously riskless bond (Bt) with a constant interest rate of r. Its instantaneous return is dBt Bt

¼ r dt:

ð3:1Þ

The second prime asset is a risky stock that represents the aggregate equity market. Its instantaneous total return dynamics are given by dSt St

¼ m dt þ

pffiffiffiffiffi Vt dZS ;

ð3:2Þ

DYNAMIC CONSUMPTION AND ASSET ALLOCATION

j

47

pffiffiffiffiffi where St denotes the price of the risky asset at time t, Vt is the time-varying instantaneous standard deviation of the return on the risky asset, and dZS is the increment of a standard Brownian motion. We assume that the short rate is constant in order to focus on the stochastic volatility of the risky asset. From these asset return dynamics, we have the assumption of constant expected excess return on the risky asset over the riskless asset, i.e. (m  r); this assumption will be relaxed in Section 3.3.2. We denote time variation with the subscript ‘t’ and let the conditional variance of the risky asset vary stochastically over time. From the following the investment opportunity is time-varying. We assume that the instantaneous variance process is  pffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffi dVt ¼ kðy  Vt Þdt þ s Vt r dZS þ 1  r2 dZV ;

ð3:3Þ

where the parameter u
0 describes the long-term mean of the variance, k  (0, 1) is the reversion parameter of the instantaneous variance process, i.e. this parameter describes the degree of mean reversion, and r is the correlation between the two Brownian motions, which is assumed to be negative to capture the asymmetric effect (Black 1976, Glosten et al. 1993). This negative correlation assumption, together with the mean-reversion of the stock return volatility, can capture major important features of the empirical literature of the equity market. In the traditional theory of derivative pricing (Black and Scholes 1973; Merton 1973b), derivative assets like options are viewed as redundant securities, the payoffs of which can be replicated by portfolios of primary assets. Thus, the market is generally assumed to be complete without the options. In this chapter we introduce derivative securities that allow the investor to include them in dynamic asset allocation strategies. If only a risky stock and a riskless bond are available for trading, then the financial market here is incomplete. The nonlinear nature of derivative securities serves to complete the market. This follows from our setting in which stock returns are not instantaneously perfectly correlated with their time-varying volatility. In this chapter the derivative securities written on the stock are non-redundant assets. In our setting derivative securities can provide differential exposure to the imperfect instantaneous correlation and make the market complete. Following Sircar and Papanicolaou (1999) and Liu and Pan (2003), in the above setting the non-redundant derivative Ot  p(St, Vt), that is the function (p) of the prices of the stock (St) and on the volatility of stock returns (Vt) at time t, will have the following price dynamics: h i pffiffiffiffiffiffiffiffiffiffiffiffiffi dOt ¼ ðm  rÞðps St þ rspv Þ þ ls 1  r2 pv þ rOt dt pffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffi pffiffiffiffiffiffiffi þ ðps St þ rspv Þ V t dZS þ ðs 1  r2 pv Þ V t dZV ;

ð3:4Þ

48 j CHAPTER 3

where l determines the stochastic volatility risk premium and ps and pv are measures of the derivative’s price sensitivity to small changes in the underlying stock price and volatility, respectively. 3.2.2

Preferences

We assume that the investor’s preferences are recursive and of the form described by Duffie and Epstein (1992a, b). Recursive utility is a generalization of the standard and time-separable power utility function that separates the elasticity of intertemporal substitution of consumption from the relative risk aversion. This means that the power utility is just a special case of the recursive utility function in which the elasticity of the intertemporal substitution parameter is the inverse of the relative risk aversion coefficient. The value function of the problem (J ) is to maximize the investor’s expected lifetime utility. We adopt the Duffie and Epstein (1992b) parameterization J ¼ Et

Z

1

f ðCt ; Jt Þdt ;

ð3:5Þ

t

where the utility f (Ct, Jt) is a normalized aggregator of an investor’s current consumption (Ct) and has the following form: 2 !1ð1=jÞ 3

1 1 C ð1  gÞJ 4 15; f ðC; J Þ ¼ b 1  1=ð1gÞ j ðð1  gÞJ Þ

ð3:6Þ

with g is the coefficient of relative risk aversion, b is the rate of time preference and 8 is the elasticity of intertemporal substitution. They are all larger than zero. The normalized aggregator f (Ct,Jt) takes the following form when 8 approaches one:  f ðC; J Þ ¼ bð1  gÞJ logðCÞ 

1 logðð1  gÞJ Þ : 1g

The investor’s objective is to maximize expected lifetime utility by choosing consumption and the proportions of wealth to invest across the risky stock and the derivative securities subject to the following intertemporal budget constraint:

  pffiffiffiffiffiffiffiffiffiffiffiffiffi pv ps St pv 2 þ ls 1  r Wt þ rWt  Ct dt þ rs dWt ¼ nt ðm  rÞWt þ pt ðm  rÞ Ot Ot Ot 

 pffiffiffiffiffiffiffiffiffiffiffiffiffi pv pffiffiffiffiffi pffiffiffiffiffi ps St pv pffiffiffiffiffi 2 Vt Wt dZV ; Vt Wt dZS þ s 1  r þ rs þnt Vt Wt dZS þ pt Ot Ot Ot 



ð3:7Þ

DYNAMIC CONSUMPTION AND ASSET ALLOCATION

j

49

where Wt represents the investor’s total wealth, nt and pt are the fractions of the investor’s financial wealth allocated to the stock and the derivative securities at time t, respectively, and Ct represents the investor’s instantaneous consumption.

3.3

OPTIMAL CONSUMPTION POLICY AND DYNAMIC ASSET ALLOCATION STRATEGIES

3.3.1

Optimal Consumption and Dynamic Asset Allocation Strategies

The principle of optimality leads to the following Bellman equation for the utility function. In the above setting, the Bellman equation becomes 

  0 ¼ sup f ðCt ; Jt Þ þ JW nt ðm  rÞWt þ pt ðm  rÞ n;p;C 

 pffiffiffiffiffiffiffiffiffiffiffiffiffi pv ps St pv 2 þls 1  r Wt þ rWt  Ct  þ rs Ot Ot O

t

 1 ps St pv 2 2 þ JV ½kðy  Vt Þ þ JWW Wt Vt ðnt Þ þ 2nt pt þ rs 2 Ot Ot 

2

2  pffiffiffiffiffiffiffiffiffiffiffiffiffi p ps St p 1 2 þ JVV s2 Vt þ JWV Wt sVt þ rs v þ s 1  r2 v þ ðpt Þ 2 Ot Ot Ot 

 

ps St p p ;  n t r þ pt þ rs v r þ sð1  r2 Þ v Ot Ot Ot

ð3:8Þ

where Jw , Jv denote the derivatives of the value function J with respect to wealth W and stochastic volatility V, respectively. We use a similar notation for higher derivatives as well. The first-order conditions for the optimization in (3.8) are Ct ¼ JWj J ð1jgÞ=ð1gÞ bj ð1  gÞ

nt ¼  þ

ð1jgÞ=ð1gÞ

;

ð3:9Þ

ðm  rÞ J  WV rs JWW Wt Vt JWW Wt JW

l ð ps St þ rspv Þ J ð ps St þ rspv Þ pffiffiffiffiffiffiffiffiffiffiffiffiffi þ WV ; pv JWW Wt Vt s 1  r2 pv JWW Wt

ð3:10Þ

l O 1 J Ot pffiffiffiffiffiffiffiffiffiffiffiffiffi t  WV : 2 JWW Wt s 1  r pv Vt JWW Wt pv

ð3:11Þ

JW

pt ¼ 

JW

Equation (3.9), which demonstrates the rule of optimal consumption, stems from the envelope condition, fc  Jw , once the value function is obtained. Equations (3.10) and (3.11) show that the optimal portfolio allocation in the stock has four components, while the derivative allocation has two. In the optimal asset allocation in both the risky stock

50 j CHAPTER 3

and the derivative, their first terms are the mean
variance portfolio weights. These are the myopie demands for an investor who only invests in a single period horizon or under a constant investment opportunity set. The second terms of both the stock and the derivative are the intertemporal hedging demands that characterize demand arising from the desire to hedge against changes in the investment opportunity set induced by the stochastic volatility. These terms are determined by the instantaneous rate of changes in relation to the value function. Without introducing the derivative security, the optimal asset allocation of the stock will contain only the first two terms. Aside from the mean
variance weights of the optimal derivative allocation, the derivative plays a role that allows the investor to insure against changes in the stochastic volatility and the time-varying investment opportunity set. The second term, i.e. the intertemporal hedging demand, of the optimal asset allocation on the stock also partially provides a similar rule, so that the introduction of the derivative makes the stock allocation interact with the derivative allocation and is expressed in the third and the fourth terms of the stock demand. The first-order conditions for our problem are not in fact explicit solutions unless we know the complicated form of the value function. After substituting the first-order conditions back into the Bellman equation and rearranging them, we conjecture that there exists a solution of the functional form J ðWt ; Vt Þ ¼ IðVt ÞðWt1g =ð1  gÞÞ. We first restrict to the special case of 8  1. We then substitute this form into Equation (3.A1) of Appendix 3.A, and the resulting ordinary differential equation will have a solution of the form I  exp(Q0  Q1Vt  Q2logVt). Rearranging this equation, we have three equations for Q2, Q1, and Q0 after collecting terms in 1/Vt , Vt , and 1. We provide the full details in Appendix 3.A. Hence, we obtain the form of the value function and the optimal consumption rule and dynamic asset allocation strategies for investing in the stock and the derivative security when 8  1. The value function is /

/

J ðWt ; Vt Þ ¼ IðVt Þ

Wt1g

¼ expðQ0 þ Q1 Vt þ Q2 log Vt Þ

1g

Wt1g 1g

:

ð3:12Þ

The investor’s optimal consumption
wealth ratio and dynamic asset allocation strategies for investing in the stock and the derivative are /

Ct ¼ b; Wt

! 1 mr 1 1 1 l ð ps St þ rspv Þ pffiffiffiffiffiffiffiffiffiffiffiffiffi Q1 þ Q2 rs  nt ¼ þ g Vt g Vt g Vt s 1  r2 pv ! 1 1 ð ps St þ rspv Þ Q1 þ Q 2  ; g Vt pv

ð3:13Þ

ð3:14Þ

DYNAMIC CONSUMPTION AND ASSET ALLOCATION

1 l 1 1 1 pffiffiffiffiffiffiffiffiffiffiffiffiffi þ Q1 þ Q2 pt ¼ 2 g Vt s 1  r ðpv =Ot Þ g Vt

!

Ot pv

j

51

ð3:15Þ

:

Equation (3.13) demonstrates the invariance of the optimal log consumption
wealth ratio to changes in volatility. It equals investors’ rate of time preference with unit elasticity of intertemporal substitution. The substitution effect and the intertemporal income effect on consumption resulting from a change in investment opportunity set perfectly cancel each other out, and the consumption of a fixed proportion of investors’ wealth each period is optimal. This is why such a consumption policy is normally referred to as being myopic (Chacko and Viceira 2005). In the more general case 8 " 1, there is no exact analytical solution. However, we can still find an approximate analytical solution following the methods described in Campbell and Viceira (2002) and Chacko and Viceira (2005). The basic idea behind the use of approximate analytical methods is that of formulating a general problem, on the condition that we can find a particular case that has a known solution, and then using that particular case and its solution as a starting point for computing approximate solutions to nearby problems. In the context of our problem, the insight we obtain is the solution for the recursive utility function when 8  1 which provides a convenient starting point for performing the expansion. Without the restriction of 8  1, the Bellman equation can be expressed as the following equation by substituting (3.9) into (3.A1) and again conjecturing that there exists a solution of the functional form J ðWt ; Vt Þ ¼ IðVt ÞðWt1g =ð1  gÞÞ: /

bj

 j 11 1  2 bI þ I ðm  rÞ þ l2 1j 2 g Vt 1j h pffiffiffiffiffiffiffiffiffiffiffiffiffii 1 1 þ IV s ðm  rÞr þ l 1  r2 þ Ir þ IV ½kðy  Vt Þ g 1g

0¼

I 1þ½ð1jÞ=ð1gÞ þ

2

1 1 1 1 ðIV Þ 2 s2 Vt þ s Vt : þ IVV 2 1g 2g I

ð3:16Þ

½ð1gÞ=ð1jÞ

To simplify, we can make the transformation IðVt Þ ¼ FðVt Þ and have a nonhomogeneous ordinary differential equation. Unfortunately, this non-homogeneous ordinary differential equation cannot be solved in closed form. Our approach is to obtain the asymptotic approximation to Equation (3.B1) shown in Appendix 3.B, by taking a log-linear expansion of the consumption
wealth ratio around its unconditional mean as introduced in the papers of Campbell and Viceira (2002) and Chacko and Viceira (2005). We provide the full details of our model’s approximate results in Appendix 3.B. We are now able to obtain the form of the value function and the optimal consumption and dynamic asset allocation strategies toward investing in the risky stock and the derivative in the stochastic environment without the constraint 8  1. The value function is /

52 j CHAPTER 3

J ðWt ; Vt Þ ¼ IðVt Þ

Wt1g 1g

¼ FðVt Þ

½ð1gÞ=ð1jÞ

Wt1g 1g

 

1g 1g ^ ^ 1 Vt þ Q ^2 log Vt Þ Wt : ðQ0 þ Q ¼ exp  1j 1g

ð3:17Þ

The investor’s optimal instantaneous consumption
wealth ratio is /

Ct Wt

^0  Q ^1 Vt  Q ^2 log Vt Þ: ¼ bj expðQ

ð3:18Þ

The investor’s optimal dynamic asset allocation strategies toward investing in the stock and the derivative are ! !

 ^1 ^2 1 1 ð ps St þ rspv Þ 1 1 Q Q ðm  rÞ  pffiffiffiffiffiffiffiffiffiffiffiffiffi l rs þ þ 1 nt ¼ g Vt g 1  j 1  j Vt s 1  r2 pv !

 ^1 ^2 1 ð ps St þ rspv Þ 1 Q Q þ ;  1 g pv 1  j 1  j Vt !

 ^1 ^ 2 1 Ot Q Q 1 l 1 1 pffiffiffiffiffiffiffiffiffiffiffiffiffi pt ¼ þ þ 1 : g s 1  r2 ðpv =Ot Þ Vt g 1  j 1  j Vt pv

ð3:19Þ

ð3:20Þ

We have now solved the approximate closed-form solution of the optimal consumption and dynamic asset allocation strategy. In the next section we provide a general extension of this model. 3.3.2

Optimal Consumption and Dynamic Asset Allocation Strategies with Time-Varying Risk Premiums

We next extend the model with constant expected excess returns of the risky stock and the constant volatility risk premiums of the derivative to a more general case in order to explore the optimal consumption rule and dynamic asset allocation strategies with timevarying instantaneous expected excess returns and time-varying stochastic volatility risk premium of the risky stock. Following Chacko and Viceira (2005), we replace the assumption of the constant excess returns on the stock with one that allows for the expected excess returns on the stock to vary with volatility:  E

!  V ¼ ðm þ m V Þdt: t 0 1 t St  r dt  dSt

ð3:21Þ

DYNAMIC CONSUMPTION AND ASSET ALLOCATION

j

53

When m1
0, the functional form implies that an increase in volatility increases risk and the expected excess return on the stock. Similarly, we replace the assumption of the constant volatility risk premium (l) with one that allows the volatility risk premium to vary with volatility in the form l0  l1Vt. Ever since the seminal work of Engle (1982), discrete-time ARCH models have become a proven approach to modeling security price volatility. For a review of the substantial literature, see Bollerslev et al. (1992). Following Nelson (1990), it is understood that GARCH-type models have well-defined continuoustime limits. Therefore, it seems reasonable to model the risk premium dependent on the conditional variance. As a basis, some empirical papers assume an ARCH/GARCH-M model with a risk premium for stochastic volatility, which is a linear function of the conditional standard deviation or variance as in this chapter. Using Heston’s (1993) option pricing formula to price currency options, Lamoureux and Lastrapes (1993) suggest a time-varying volatility risk premium. Empirical applications of the ARCH-M model have met with much success. Some studies (see, e.g. French et al. (1987), Chou (1988) and Campbell and Hentschel (1992)) have reported consistently positive and significant estimates of the risk premium. When l1
0, the functional form implies that an increase in volatility increases risk and also the stochastic volatility risk premium. When m1  0 and l1  0, the results of this section will reduce to those of Section 3.3.1. In the above setting, the non-redundant derivative (Ot  p(St, Vt)), which is a function (p) of the prices of the stock (St) and the volatility of stock returns (Vt) at time t, will have the following price dynamics: i h pffiffiffiffiffiffiffiffiffiffiffiffiffi dOt ¼ ðm0 þ m1 Vt Þðps St þ rspv Þ þ ðl0 þ l1 Vt Þs 1  r2 pv þ rOt dt pffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffi pffiffiffiffiffi þ ðps St þ rspv Þ Vt dZS þ ðs 1  r2 pv Þ Vt dZV :

ð3:22Þ

From these extensions, the Bellman equation can be expressed as

bj

 j 11 1  bI þ I ðm0 þ m1 Vt Þ2 þ ðl0 þ l1 Vt Þ2 1j 2 g Vt 1j h pffiffiffiffiffiffiffiffiffiffiffiffiffii 1 þ Ir þ IV s ðm0 þ m1 Vt Þr þ ðl0 þ l1 Vt Þ 1  r2 g

0¼

I 1þ½ð1jÞ=ð1gÞ þ

1 1 1 1 1 ðIV Þ2 2 2 s Vt þ s Vt : ½kðy  Vt Þ þ IVV þ IV 1g 2 1g 2g I

ð3:23Þ

We follow the same method in Section 3.3.1 to derive the optimal consumption and dynamic asset allocation strategies under these assumptions. The main steps of the derivation are guessing the same functional form for J(Wt ,Vt) and I(Vt) as in Appendix 3.B. We then can reduce the Bellman equation to an ODE in F(Vt) as follows:

54 j CHAPTER 3

j

0 ¼ b F

1

" # 1 2 1 2 2 2 þ jb þ ð1  jÞ  ðm0 þ l0 Þ þ 2ðm0 m1 þ l0 l1 Þ þ ðm1 þ l1 ÞVt 2g Vt

pffiffiffiffiffiffiffiffiffiffiffiffiffi  pffiffiffiffiffiffiffiffiffiffiffiffiffi i 1  g FV h s m0 r þ l0 1  r2 þ m1 r þ l1 1  r2 Vt g F "

#  2 FV 1 1g FV FVV 2 þ1 kðy  Vt Þ þ s Vt þ ð1  jÞr   2 1j F F F



2 1 ð1  gÞ2 1 FV þ s 2 Vt : 2 1j F g

ð3:24Þ

Following the steps in Equation (3.A8) to Equation (3.A10), we obtain the solution of the ~0 þ Q ~ 1 Vt þ Q ~2 log Vt Þ, and can express Equation ODE in F(Vt) with FðVt Þ ¼ expðQ (3.24) as

  1 ~ ~ ~ 0 ¼  f0 þ f1 j log b  Q0  Q1 Vt  Q2 log y þ Vt  1 y  1 2 1 2 2 2 þ jb þ ð1  jÞ  ðm0 þ l0 Þ þ 2ðm0 m1 þ l0 l1 Þ þ ðm1 þ l1 ÞVt 2g Vt

 

 pffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffi 1g ~ 1 2 2 ~ s  m0 r þ l0 1  r þ m1 r þ l1 1  r Vt  Q1 þ Q2 g Vt 



 2 1 1 1  g 1 ~1 þ Q ~1 þ Q ~2 ~2 þ1 Q þ ð1  jÞr  Q kðy  Vt Þ þ Vt 2 1j Vt

 

1 2 ~ 1 1 ð1  gÞ2 1 1 2 2 2 ~ ~ ~ ~ s Vt þ  Q1 þ Q2 þ Q2 s Vt : ð3:25Þ Q1 þ Q2 Vt Vt2 2 1j Vt g 



Rearranging this equation after collecting terms in 1/Vt , Vt and 1 we have three equations ~1 , and Q ~0 : ~2, Q for Q "

1 1g

1 ð1  gÞ2

1

#

~22 s2 þ s2 Q 2 1j 2 1j g

 1g 1  g pffiffiffiffiffiffiffiffiffiffiffiffiffi2 1 2 ~ sm0 r þ l0 1  r s þ ky  s Q2  g g 2 þ

  1 ð1  jÞ m20 þ l20 ¼ 0; 2g

ð3:26Þ

DYNAMIC CONSUMPTION AND ASSET ALLOCATION

j

55

"

#  2 pffiffiffiffiffiffiffiffiffiffiffiffiffi 1 1  g 2 1 ð1  gÞ 1 1g  2 ~2 2 ~1 s þ s Q1 þ f 1  s m1 r þ l 1 1  r þ k Q 2 1j 2 1j g g   1 ~ 1 ð1  jÞ m21 þ l21 ¼ 0; ð3:27Þ þ f1 Q 2 þ y 2g



pffiffiffiffiffiffiffiffiffiffiffiffiffi 1g  ~2 s m1 r þ l1 1  r2 þ k Q g " #

 2 1g 1  g pffiffiffiffiffiffiffiffiffiffiffiffiffi2 1  g ð1  gÞ 1 ~2 ~1 Q ~1 þ sm0 r þ l0 1  r s þ ky Q s2 þ s2 Q  g g 1j 1j g ~0  f0  f1 j log b þ jb þ 1 ð1  jÞðm0 m1 þ l0 l1 Þ þ ð1  jÞr ¼ 0: þ f1 Q ð3:28Þ g

f1 log y  f1 

We are now able to obtain the value function and the investor’s optimal consumption and dynamic asset allocation strategy for investing in the stock and derivative security with time-varying risk premiums. The value function is J ðWt ; Vt Þ ¼ IðVt Þ

Wt1g

¼ FðVt Þ½ð1gÞ=ð1jÞ

Wt1g

1g 1g  

 Wt1g 1g ~ ~ ~ : ¼ exp  Q0 þ Q1 Vt þ Q2 log Vt 1j 1g

ð3:29Þ

The investor’s optimal instantaneous consumption
wealth ratio is /

Ct Wt

  ~0  Q ~1 Vt  Q ~2 log Vt : ¼ bj exp Q

ð3:30Þ

The investor’s optimal dynamic asset allocation strategy toward investing in the stock and the derivative security is !  ~1 ~2 1 Q Q 1 þ 1 rs þ g 1  j 1  j Vt ! 1 1 ð ps St þ rspv Þ pffiffiffiffiffiffiffiffiffiffiffiffiffi l1 þ l0  g Vt s 1  r 2 p v !

 ~1 ~2 1 ð ps St þ rspv Þ Q Q 1 þ  1 g pv 1  j 1  j Vt ! ! 1 1 1 1 r pS pffiffiffiffiffiffiffiffiffiffiffiffiffi  pt s t ; m1 þ m0 l1 þ l0  ¼ 2 g Vt g Vt Ot 1r

1 1 m1 þ m0 nt ¼ g Vt

!



ð3:31Þ

56 j CHAPTER 3

#  ~ 

~ 2 1  Ot 1 1 1 1 Q1 Q pffiffiffiffiffiffiffiffiffiffiffiffiffi l1 þ l0 þ 1 þ pt ¼ : g s 1  r2 Vt g 1  j 1  j Vt pv 

ð3:32Þ

We have solved for the approximate solution of the optimal consumption and dynamic asset allocation strategies with time-varying risk premiums. In the next section we discuss the optimal consumption rule and dynamic asset allocation of our results and the effects of the introduction of derivative securities. We also provide analyses of our results, in this section and of a numerical example shown in the figures.

3.4

ANALYSES OF THE OPTIMAL CONSUMPTION RULE AND DYNAMIC ASSET ALLOCATION STRATEGY

In Section 3.3.2 we have the most general results. Equation (3.30) shows that the optimal log consumption
wealth ratio is a function of stochastic volatility, with coefficients ~1 =ð1  jÞ and Q ~2 =ð1  jÞ. While Q ~2 is the solution to the quadratic Equation (3.26), Q ~ ~2 , and Q ~0 is the solution to Equation (3.28) Q1 is the solution to Equation (3.27) given Q ~2 . When g
1 for coefficient Q ~2 , Equation (3.26) has two real roots of ~1 and Q given Q opposite signs. In each quadratic equation, we would like to know which solution is good for our problem from the following criteria. First, we must ensure that the roots of the discriminant are real. We must then determine the sign of the roots that we will choose. Campbell and Viceira (1999, 2002), Campbell et al. (2004) and Chacko and Viceira (2005) show that only one of them maximizes the value function. This is also the only root that ~2 is equal to zero when g  8  1, that is, in the log utility case. Under these ensures Q criteria, the value function J is maximized only with the solution associated with the negative root of the discriminant of the quadratic Equation (3.26), i.e. the positive root of ~2 =ð1  jÞ > 0. Equation (3.26). It can immediately be shown that Q ~1 =ð1  jÞ < 0. Thus, the ratio By the same criteria, it can immediately be shown that Q of consumption to wealth is shown to be an increasing function of volatility for those investors whose elasticity of intertemporal substitution of consumption (8) is less than one. Conversely, this ratio is a decreasing function of volatility when the elasticity of intertemporal substitution of consumption for an investor is greater than one. (It is ~2 =Vt > 0.) The relative importance of intertemporal income and ~1  Q assumed that Q substitution effects of volatility on consumption is thus demonstrated. For illustration, suppose the greater effect of volatility on consumption. The increase in volatility provides greater opportunities for investment, because derivative securities can provide volatility exposures. Moreover, assuming positively priced volatility risk, returns on the portfolio do not increase in volatility, yet there is an improvement in the expected return. A negative intertemporal substitution effect on consumption will result from these investment opportunities, since the latter are more favorable than at other times. A positive income effect also results, as an increase in expected returns decreases the marginal utility of consumption. The income effect dominates the substitution effect for investors whose /

DYNAMIC CONSUMPTION AND ASSET ALLOCATION

j

57

8 B 1 and their current consumption rises relative to wealth. Investors whose 8
1 will see the substitution effect dominate and will therefore cut their current consumption relative to wealth. The optimal dynamic asset allocation strategy for the risky stock has four components, as with the first equality of Equation (3.31) and in Figure 3.1. If we do not introduce any derivative security and the investor holds only the risky stock, then the optimal dynamic asset allocation for the risky stock will only have the first two components in the first equality of Equation (3.31), i.e. the myopic component and the intertemporal hedging component. First, the dependence of the myopic component is simple. It is an affine function of the reciprocal of the time-varying volatility and decreases with the coefficient of relative risk aversion. Since volatility is time-varying, the myopic component is timevarying too. The position of the myopic component can be either positive or negative, depending on m0, m1, and the level of volatility. We know that if m1  0 and m0  m  r, then the result also nests the model results in Section 3.3.1. An extension of the replacement of the constant expected excess return with one that allows an expected excess return on the risky stock to vary with volatility implies that increased risk is rewarded with an increase in expected excess return when m1
0. Hence, the investor will take a long position on the myopic component of the risky stock. The intertemporal hedging component of the optimal dynamic asset allocation for the risky stock is an affine function of the reciprocal of the time-varying volatility, with 0.90

0.70

0.50 0.30

0.10

–0.10

1.60

2.20

2.80

3.40

4.00

4.60 Gamma

5.20

5.80

6.40

7.00

7.60

–0.30 –0.50 The optimal dynamic asset allocation for the risky stock The myopic component The intertemporal hedging component The myopic correction term for holding derivatives The intertemporal correction term for holding derivatives

FIGURE 3.1 The optimal dynamic asset allocation strategy toward investing in the stock and its components in relation to g.

58 j CHAPTER 3

~1 =ð1  jÞ and Q ~2 =ð1  jÞ. Since Q ~2 =ð1  jÞ > 0, the sign of the coefficient coefficients Q of the intertemporal hedging demand coming from pure changes in time-varying volatility is positive when g
1 [this is true only for r
0]. If we do not introduce any derivative security and instead hold only risky stock, then the intertemporal hedging component for the risky stock will consist of the correlation or asymmetric effect. The intertemporal hedging component of the optimal asset allocation for risky stock is affected by the instantaneous correlation between the two Brownian motions. If the instantaneous correlation is perfect, then markets are complete, without the need for holding any derivative securities. However, we allow for imperfect instantaneous correlation in the model. In particular, if r B 0, this means that the unexpected return of the risky asset is low (the market situation is bad), and then the state of the market uncertainty will be ^2 =ð1  jÞ > 0 when g
1, the negative instantaneous correlation implies the high. Since Q investor will have negative intertemporal hedging demand due to changes solely in the volatility of the risky asset, which lacks the hedging ability against an increase in volatility. Similar discussions are found in Liu (2005) and Chacko and Viceira (2005). However, in our generalized model the consideration of holding derivative securities complicates the asset allocation strategies for long-horizon investors. The other two components of the optimal dynamic asset allocation for the risky stock are correction terms for holding the derivative. The first and the second component of these two correction terms (i.e. the third and the fourth terms of the optimal dynamic asset allocation for the risky stock) are the myopic correction term and the intertemporal correction term for the derivatives held, respectively. These terms are from the interaction between the derivative and its underlying stock. We can see that from the first equality to the second equality of Equation (3.31), the intertemporal hedging demand on the risky stock will be canceled by the correction term of holding the derivative. In the second equality of Equation (3.31), in its first term we show that the net demand for the risky stock will finally link to the risk-and-return tradeoff associated with price risk, because volatility exposures have been captured by holding the derivatives. The second term is the correction term of the correlation effect, and the third correction term is to correct for the delta effect of the derivative held. The relationships between these components with the degree of risk aversion (g) are also seen in Figure 3.2. The optimal dynamic asset allocation for the derivative depends on the proportion of derivative’s price to its volatility exposure. This proportion measures the dollars expended on the derivative for each unit of volatility exposure. The less expenditure that a derivative provides for the same unit of volatility exposure, the more effective it is as a vehicle to hedge volatility risk. Therefore, a smaller portion of the investor’s wealth need be allocated to the derivative (Liu and Pan 2003). Furthermore, the optimal dynamic asset allocation strategy for the derivative security can also be separated into two components: the myopic demand and the intertemporal hedging demand, which are also seen in Figure 3.3. The myopic demand is an affine function of the reciprocal of the time-varying volatility with coefficients of l0 and l1, and it depends on (1/g). The time-varying volatility also makes the myopic demand time-varying. The long or short position of the myopic derivative’s

j

DYNAMIC CONSUMPTION AND ASSET ALLOCATION

59

0.60

0.50

0.40

0.30

0.20

0.10

0 1.60

2.20

2.80

3.40

4.00 4.60 Gamma

5.20

5.80

6.40

7.00

7.60

The optimal dynamic asset allocation for the risky stock The myopic component The correction term of the correlation effect The correction term of the delta effect

FIGURE 3.2 The optimal dynamic asset allocation strategy toward investing in the risky stock and its components in relation to g.

0.0112 0.39 0.0110 0.34 0.0108

0.0106

0.29

0.0104 0.24 0.0102

0.0100

0.19 1.50 2.00 2.50 3.00 3.50 4.00 4.50 5.00 5.50 6.00 6.50 7.00 7.50 Gamma The optimal dynamic asset allocation for the derivative security The intertemporal hedging demand The myopic demand

FIGURE 3.3 The optimal dynamic asset allocation strategy toward investing in the derivative security and its components in relation to g.

60 j CHAPTER 3

demand depends on the volatility
risk premium parameters l0 and l1, and the level of volatility. The second term of the optimal dynamic asset allocation of the derivative is the intertemporal hedging component, which depends on all the parameters that characterize investor preferences and the investment opportunity set. Without loss of generality we may assume that the option is volatility exposure is positive. In our setting, the instantaneous correlation between unexpected returns of the derivative and changes in ~2 =ð1  jÞ > 0 when g
1, for more riskvolatility is perfect positive. In addition, since Q averse investors the intertemporal hedging demand is positive due to changes solely in the stochastic volatility. Investors who are more risk-averse than logarithmic investors have a positive intertemporal hedging demand for the derivative, because it tends to do better when there is an increase in volatility risk. The derivative provides hedging ability against an increase in volatility. /

3.5

CONCLUSIONS

Ever since Merton (1969, 1971) introduced the standard intertemporal consumption and investment model, it has been studied extensively in the finance literature and has become a classical problem in financial economics. The literature on the broad set of issues of intertemporal consumption and investment or optimal dynamic asset allocation strategies deals with investors’ access only to bond and stock markets and excludes the derivatives market. While a few recent studies include derivative securities in the investment portfolio, investors are assumed to have a specified utility defined over wealth at a single terminal date, abstraction from the choice of consumption over time, and the studies are restricted to a static position in derivative securities or the construction of a buy-andhold portfolio. This chapter considers a model in which a long-term investor chooses consumption as well as optimal dynamic asset allocation including a riskless bond, risky stock and derivatives on the stock when there is predictable variation in return volatility. We then maximize a more general recursive utility function defined over intermediate consumption rather than terminal wealth to reflect the realistic problem facing an investor saving for the future. We show that the optimal log consumption
wealth ratio is a function of stochastic volatility. Furthermore, the consumption
wealth ratio is an increasing function of volatility for investors whose elasticity of intertemporal substitution of consumption is smaller than one, while it is a decreasing function of volatility for investors whose elasticity of intertemporal substitution of consumption is larger than one. This result reflects the comparative importance of intertemporal income and substitution effects of volatility on consumption. Merton (1971, 1973a) shows that dynamic hedging is necessary for forward-looking investors when investment opportunities are time-varying. In this chapter we show that considering derivative securities in portfolio decisions to create a dynamic asset allocation strategy brings benefits of improvements to the hedging ability in the intertemporal hedging component. When we introduce a non-redundant derivative written on the risky stock in the incomplete financial market in this optimal dynamic asset allocation, the /

/

DYNAMIC CONSUMPTION AND ASSET ALLOCATION

j

61

derivative in the asset allocation can provide differential exposures to stochastic volatility and make the market complete. Derivative securities are a significant tool for expanding investors’ dimensions for risk-and-return tradeoffs, as a vehicle to hedge the additional risk factor of stochastic volatility in the stock market. Non-myopic investors utilize derivative securities, which provide access to volatility risk, to capitalize upon the timevarying nature of their opportunity set.

REFERENCES Bhamra, H. and Uppal, R., The role of risk aversion and intertemporal substitution in dynamic consumption-portfolio choice with recursive utility. EFA Annual Conference Paper No. 267, 2003. Black, F., Studies of stock market volatility changes, in Proceedings of the American Statistical Association, Business and Economic Statistics Section, 1976, pp. 177
/181. Black, F. and Scholes, M.S., The pricing of options and corporate liabilities. J. Polit. Econ., 1973, 81, 637
/654. Bollerslev, T., Chou, R.Y. and Kroner, K.F., ARCH modeling in finance: A review of the theory and empirical evidence. J. Econometrics, 1992, 52, 5
/59. Campbell, J.Y., Intertemporal asset pricing without consumption data. Am. Econ. Rev., 1993, 83, 487
/512. Campbell, J.Y. and Hentschel, L., No news is good news: An asymmetric model of changing volatility in stock returns. J. Finan. Econ., 1992, 31, 281
/318. Campbell, J.Y. and Viceira, L.M., Consumption and portfolio decisions when expected returns are time varying. Q. J. Econ., 1999, 114, 433
/495. Campbell, J.Y. and Viceira, L.M., Who should buy long-term bonds? Am. Econ. Rev., 2001, 91, 99
/127. Campbell, J.Y. and Viceira, L.M., Strategic Asset Allocation: Portfolio Choice for Long-term Investors, 2002 (Oxford University Press: Oxford). Campbell, J.Y., Chacko, G., Rodriguez, J. and Viceira, L.M., Strategic asset allocation in a continuous-time VAR model. J. Econ. Dynam. Contr., 2004, 28, 2195
/2214. Chacko, G. and Viceira, L., Dynamic consumption and portfolio choice with stochastic volatility in incomplete markets. Rev. Finan. Stud., 2005, 18, 1369
/1402. Chalasani, P. and Jha, S., Randomized stopping times and American option pricing with proportional transaction cost. Math. Finan., 2001, 11, 33
/77. Chou, R.Y., Volatility persistence and stock valuations: Some empirical evidence using GARCH. J. Appl. Econometrics, 1988, 3, 279
/294. Cont, R., Tankov, P. and Voltchkova, E., Hedging with options in models with jumps, in Proceedings of the II Abel Symposium 2005 on Stochastic Analysis and Applications, 2005. Duffie, D. and Epstein, L.G., Stochastic differential utility. Econometrica, 1992a, 60, 353
/394. Duffie, D. and Epstein, L.G., Asset pricing with stochastic differential utility. Rev. Finan. Stud., 1992b, 5, 411
/436. Engle, R.F., Autoregressive conditional heteroscedasticity with estimates of the variance of UK inflation. Econometrica, 1982, 50, 987
/1008. French, K.R., Schwert, G.W. and Stambaugh, R.F., Expected stock returns and volatility. J. Finan. Econ., 1987, 19, 3
/30. Glosten, L.R., Jagannathan, R. and Runkle, D., On the relation between the expected value and the volatility of the nominal excess return on stocks. J. Finan., 1993, 48, 1779
/1801. Haugh, M. and Lo, A., Asset allocation and derivatives. Quant. Finan., 2001, 1, 45
/72.

62 j CHAPTER 3

Heston, S.L., A closed-form solution for options with stochastic volatility with applications to bond and currency options. Rev. Finan. Stud., 1993, 6, 327
/343. Ilhan, A., Jonsson, M. and Sircar, R., Optimal investment with derivative securities. Finan. Stochast., 2005, 9, 585
/595. Lamoureux, C.G. and Lastrapes, W.D., Forecasting stock-return variances: Toward an understanding of stochastic implied volatilities. Rev. Finan. Stud., 1993, 6, 293
/326. Liu, J., Portfolio selection in stochastic environments. Stanford GSB Working Paper, 2005. Liu, J. and Pan, J., Dynamic derivative strategies. J. Finan. Econ., 2003, 69, 401
/430. Merton, R.C., Lifetime portfolio selection under uncertainty: The continuous time case. Rev. Econ. Statist., 1969, 51, 247
/257. Merton, R.C., Optimum consumption and portfolio rules in a continuous-time model. J. Econ. Theor., 1971, 3, 373
/413. Merton, R.C., An intertemporal capital asset pricing model. Econometrica, 1973a, 41, 867
/887. Merton, R.C., The theory of rational option pricing. Bell J. Econ. Mgmt. Sci., 1973b, 4, 141
/183. Nelson, D., ARCH models as diffusion approximations. J. Econometrics, 1990, 45, 7
/38. Sircar, K.R. and Papanicolaou, G.C., Stochastic volatility, smile & asymptotics. Appl. Math. Finan., 1999, 6, 107
/145. Stein, E.M. and Stein, J., Stock price distributions with stochastic volatility: An analytic approach Rev. Finan. Stud., 1991, 4, 727
/752.

APPENDIX 3.A: DERIVATION OF THE EXACT SOLUTION WHEN j  1 Substituting the first-order conditions (i.e. Equations (3.9)
(3.11)) back into the Bellman equation (i.e. Equation (3.8)) and rearranging we get /

2  1 ðJW Þ 1  ðm  rÞ2 þ l2 0 ¼ f ðCðJ Þ; J Þ  JW CðJ Þ  2 JWW Vt h pffiffiffiffiffiffiffiffiffiffiffiffiffii   J J  W WV s ðm  rÞr þ l 1  r2 þ JW rWt þ JV kðy  Vt Þ JWW

1 1 ðJWV Þ2 2 s Vt : þ JVV s2 Vt  2 2 JWW

ð3:A1Þ

We conjecture that there exists a solution of the functional form J ðWt ; Vt Þ ¼ IðVt Þ  ½Wt1g =ð1  gÞ when 8  1, and substituting it into Equation (3.A1) we obtain

0¼

log b 

1 1g

 log I  1 bI þ

 1  2 ðm  rÞ þ l2 2 g Vt 11

I

pffiffiffiffiffiffiffiffiffiffiffiffiffii  1  kðy  Vt Þ þ IV s ðm  rÞr þ l 1  r2 þ Ir þ IV g 1g 1

h

1 1 1 1 ðIV Þ2 2 s 2 Vt þ s Vt : þ IVV 2 1g 2g I The above ordinary differential equation has a solution of the I ¼ expðQ0 þ Q1 Vt þ Q2 log Vt , and so Equation (3.A2) can be expressed as

ð3:A2Þ form

DYNAMIC CONSUMPTION AND ASSET ALLOCATION

 0¼

j

63



  1 1 Q0 þ Q1 Vt þ Q2 log y þ Vt  1  1 b log b  1g y ! h pffiffiffiffiffiffiffiffiffiffiffiffiffii  1 11 1  1 2 2 Q1 þ Q 2 s ðm  rÞr þ l 1  r2 ðm  rÞ þ l þ þ 2 g Vt g Vt þr þ

1 Q1 þ Q 2 Vt

1 1 þ 2 1g

"

!

1 ½kðy  Vt Þ 1g

1 Q 1 þ Q2 Vt

!2

# !2 1 2 11 1 Q1 þ Q 2 s Vt þ  Q2 s 2 Vt : Vt2 2g Vt

ð3:A3Þ

Rearranging that equation, we have three equations for Q2, Q1, and Q0 after collecting terms in 1/Vt , Vt , and 1:

  1 1 11 2 1 1 pffiffiffiffiffiffiffiffiffiffiffiffiffi2 1 1 1 2 2 2 s þ s Q2 þ sðm  rÞr þ sl 1  r þ ky  s Q2 2 1g 2g g g 1g 2 1g þ





 1 1 2 ðm  rÞ þ l2 ¼ 0; 2g

ð3:A4Þ

  1 1 11 2 1 1 1 1 2 2 s þ s Q1  bþ k Q1  b Q2 ¼ 0; 2 1g 2g 1g 1g 1g y

ð3:A5Þ

  1 1 1 1 1 pffiffiffiffiffiffiffiffiffiffiffiffiffi2 1 b log y þ b k Q2 þ sðm  rÞr þ sl 1  r þ ky Q1  1g 1g 1g g g 1g  1 1 2 1 2 s þ s Q1 Q2  bQ0 þ b log b  b þ r ¼ 0; þ ð3:A6Þ 1g g 1g

and in Equation (3.A4), we have

Q2 ¼

b

pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi b2  4ac 2a

;

64 j CHAPTER 3

where a¼

1 1 11 2 s2 þ s; 2 1g 2g

1 b ¼ sðm  rÞr g

1 pffiffiffiffiffiffiffiffiffiffiffiffiffi 1 1 1 ky  s2 ; þ sl 1  r2 þ g 1g 2 1g and c¼

 1 1 ðm  rÞ2 þ l2 : 2g

APPENDIX 3.B: DERIVATION OF THE APPROXIMATE RESULTS For simplicity, we can make the transformation IðVt Þ ¼ FðVt Þ following non-homogeneous ordinary differential equation: 0 ¼ bj F1 þ jb þ

½ð1gÞ=ð1jÞ

and obtain the

  1 1 2 ð1  jÞ ðm  rÞ þ l2  2g Vt

pffiffiffiffiffiffiffiffiffiffiffiffiffii 1  g FV h F s ðm  rÞr þ l 1  r2 þ ð1  jÞr  V kðy  Vt Þ g F F "

#  2

2 1 1g FV FVV 2 1 ðg  1Þ2 1 FV þ1 þ s Vt þ  s2 Vt : 2 1j 2 1j F F F g



ð3:B1Þ

From the transformation, we obtain the envelope condition of Equation (3.9):1 Ct Wt

( j

¼b F

1

¼ exp log

Ct Wt

!) expfct  wt g:

ð3:B2Þ

From Equation (3.9) we have Ct ¼ JWj J ð1jgÞ=ð1gÞ bj ð1  gÞð1jgÞ=ð1gÞ , and we also conjecture that there exists a solution of the functional form J ðWt ; Vt Þ ¼ IðVt Þ½Wt1g =ð1  gÞ , and we make the transformation ð1gÞ=ð1jÞ IðVt Þ ¼ FðVt Þ . We thus have J ¼ ½Wt1g =ð1  gÞ Fð1gÞ=ð1jÞ and JW ¼ Fð1gÞ=ð1jÞ Wtg . Therefore, we have 1

Ct ¼ JWj J ð1jgÞ=ð1gÞ bj ð1  gÞð1jgÞ=ð1gÞ

1g ð1jgÞ=ð1gÞ j Wt ð1jgÞ=ð1gÞ Fð1gÞ=ð1jÞ ¼ ðFð1gÞ=ð1jÞ Wt g Þ bj ð1  gÞ 1g ¼ F1 bj Wt :

DYNAMIC CONSUMPTION AND ASSET ALLOCATION

j

65

Using a first-order Taylor expansion of exp{ct  wt} around the expectation of (ct  wt), we can write bj F1  expfEðct  wt Þg þ expfEðct  wt Þg  ½ðct  wt Þ  Eðct  wt Þ ¼ expfEðct  wt Þg  f1  Eðct  wt Þg þ expfEðct  wt Þg  ðct  wt Þ f0 þ f1 ðct  wt Þ:

ð3:B3Þ

Substituting (3.B3) into Equation (3.B1) and assuming this equation has a solution of the ^0 þ Q ^ 1 Vt þ Q ^2 log Vt Þ, from this guessed solution an Equation (3.B2) form FðVt Þ ¼ expðQ we can show that n   o ^0 þ Q ^ 1 Vt þ Q ^2 log Vt Þ 1 ðct  wt Þ ¼ log bj expðQ ^ 1 Vt  Q ^2 log Vt : ^0  Q ¼ j log b  Q

ð3:B4Þ

As such, we can express Equation (3.B1) as 

  1 ^ ^ ^ 0 ¼  f0 þ f1 j log b  Q0  Q1 Vt  Q2 log y þ Vt  1 y

 h h i pffiffiffiffiffiffiffiffiffiffiffiffiffii 1 1g ^ 1 2 2 1 ^ Q1 þ Q2 s ðm  r Þr þ l 1  r2 þ jb þ ð1  jÞ ðm  r Þ þl  2g Vt g Vt ! 1 ^1 þ Q ^2 kðy  Vt Þ þ ð1  jÞr  Q Vt 

1 þ 2



# 

2

2 1g 1 1 1 ^2 ^2 ^2 ^1 þ Q ^1 þ Q þ1 Q s 2 Vt  Q þQ 1j Vt Vt Vt2

2 2 1 ð1  gÞ 1 1 1 ^2 ^ þQ þ s 2 Vt : Q t 2 1j V g

ð3:B5Þ

^1 , and Q ^0 : ^2, Q Rearranging the above equation we have the following three equations for Q "

# 1 1  g 2 1 ð1  gÞ2 1 ^22 s þ s2 Q 2 1j 2 1j g  1g 1  g pffiffiffiffiffiffiffiffiffiffiffiffiffi2 1 2 ^ sðm  rÞr þ l 1  r s þ ky s Q2  g g 2 þ

1 2 ð1  jÞ½ðm  rÞ þ l2 ¼ 0; 2g

ð3:B6Þ

66 j CHAPTER 3

"

# 2 1 1  g 2 1 ð1  gÞ 1 2 ^2 ^2 ¼ 0; ^ 1 þ f1 1 Q s þ s Q1 þ ðf1 þ kÞQ 2 1j 2 1j y g

ð3:B7Þ

^2 ðf1 log y  f1 þ kÞQ  1g 1  g pffiffiffiffiffiffiffiffiffiffiffiffiffi2 ^1 sðm  rÞr þ l 1  r s þ ky Q  g g "

# 1  g 2 ð1  gÞ2 1 ^ 2 þ f1 Q ^ 0  f0 ^1 Q s þ s2 Q þ 1j 1j g  f1 j log b þ jb þ ð1  jÞr ¼ 0;

ð3:B8Þ

^1 can be found from ^2 can be found from the quadratic Equation (3.B6), Q where Q ^0 can be found from Equation (3.B8), given Q ^1 and Q ^2 . ^2 , and Q Equation (3.B7) given Q

4

CHAPTER

Volatility-Induced Financial Growth M. A. H. DEMPSTER, IGOR V. EVSTIGNEEV and KLAUS ´ R. SCHENK-HOPPE CONTENTS 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 The Model and the Main Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Proofs of the Main Results and Their Explanation . . . . . . . . . . . . . . . . 4.4 Stationary Markets: Puzzles and Misconceptions. . . . . . . . . . . . . . . . . 4.5 Growth Acceleration, Volatility Reduction and Noise-Induced Stability . 4.6 Volatility-Induced Growth under Small Transaction Costs . . . . . . . . . . 4.7 Growth under Transaction Costs: An Example . . . . . . . . . . . . . . . . . . 4.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4.1

. . . . . . . . . .

. . . . . . . . . .

67 69 72 74 76 79 81 83 83 83

INTRODUCTION

C

AN VOLATILITY, WHICH IS PRESENT in virtually every financial market and usually thought of in terms of a risky investment’s downside, serve as an ‘engine’ for financial growth? Paradoxically, the answer to this question turns out to be positive. To demonstrate this paradox, we examine the long-run performance of constant proportions investment strategies in a securities market. Such strategies prescribe rebalancing the investor’s portfolio, depending on price fluctuations, so as to keep fixed proportions of wealth in all the portfolio positions. Assume that asset returns form a stationary ergodic process and asset prices grow (or decrease) at a common asymptotic rate r. It is shown in this chapter that if an investor employs any constant proportions strategy, then the value of his/her portfolio grows almost surely at a rate strictly greater than r, provided that the investment proportions are strictly positive and the stochastic price process is in a sense non-degenerate. The very mild assumption of non-degeneracy

67

68 j CHAPTER 4

we impose requires at least some randomness, or volatility, of the price process. If this assumption is violated, then the market is essentially deterministic and the result ceases to hold. Thus, in the present context, the price volatility may be viewed as an endogenous source of acceleration of financial growth. This phenomenon might seem counterintuitive, especially in stationary markets (Evstigneev and Schenk-Hoppe´ 2002; Dempster et al. 2003), where the asset prices themselves, and not only their returns, are stationary. In this case, r
0, i.e. each asset grows at zero rate, while any constant proportions strategy exhibits growth at a strictly positive exponential rate with probability one! To begin with, we focus on the case where all the assets have the same growth rate r. The results are then extended to a model with different growth rates r 1 ; . . . ; r K . In this setting, a constant proportions strategy with proportions l 1 > 0; . . . ; l K > 0 grows almost P surely at a rate strictly greater than k l k r k (see Theorem 4.1 in Section 4.2). The phenomenon highlighted in this paper has been mentioned in the literature, but its analysis has been restricted to examples involving specialized models. The terms ‘excess growth’ (Fernholz and Shay 1982) and the more discriptive ‘volatility pumping’ (Luenberger 1998) have been used to name similar effects to those discussed here. Cover (1991) used the mechanism of volatility pumping in the construction of universal portfolios. These ideas have been discussed in connection with financial market data in Mulvey and Ziemba (1998), Mulvey (2001) and Dries et al. (2002). Such questions have typically been studied in the context of maximization of expected logarithmic utilities— ‘log-optimal investments’ (Kelly 1956; Breiman 1961; Algoet and Cover 1988; MacLean et al. 1992; Hakansson and Ziemba 1995; Li 1998; Aurell et al. 2000). In this chapter we ignore questions of optimality of trading strategies and do not use the related notion of expected utility involved in optimality criteria.1 Constant proportions strategies play an important role in various practical financial computations, see e.g. Perold and Sharpe (1995). The assumption of stationarity of asset returns is widely accepted in financial theory and practice allowing, as it does, expected exponential price growth and mean reversion, volatility clustering and very general intertemporal dependence, such as long memory effects, of returns. However, no general results justifying and explaining the fact of volatility-induced growth have been established up to now. In spite of the fundamental importance and generality of this fact, no results pertaining to an arbitrary constant proportions strategy (regardless of its optimality) and any securities market with stationary non-degenerate asset returns have been available in the literature. The purpose of this chapter is to fill this gap. Most of our results are rather easy consequences of some general mathematical facts, and the mathematical aspects do not play a crucial role. The main contribution of the present work is that we pose and analyse a number of questions that have not been systematically analysed before. These questions are especially interesting because the common intuition 1

In connection with the discussion of relevant literature, we can mention a strand of publications dealing with Parrondo games (Harmer and Abbott 1999). Models considered in those publications are based on the analysis of lotteries whose odds depend on the investor’s wealth. It is pointed out that losing lotteries, being played in a randomized alternating order, can become winning. In spite of some similarity, there are no obvious direct links between this phenomenon and that studied in the present chapter.

VOLATILITY-INDUCED FINANCIAL GROWTH

j

69

currently prevailing in the mathematical finance community suggests wrong answers to them (see the discussion in Section 4.4). Therefore it is important to clarify the picture in order to reveal and correct misconceptions. This is a central goal in this study. The chapter is organized as follows. In Section 4.2 we describe the model, formulate the assumptions and state the main results. Section 4.3 contains proofs of the results and a discussion of their intuitive meaning. In Sections 4.4 and 4.5 we analyse the phenomenon of volatility-induced growth from various angles, focusing primarily on the case of stationary prices. We answer a number of questions arising naturally in connection with the theory developed. In Section 4.6, we show how this theory can be extended to markets with small transaction costs. Section 4.7 analyses an example in which estimates for the size of transaction cost rates allowing volatility-induced growth can be established.

4.2

THE MODEL AND THE MAIN RESULTS

Consider a financial market with K ] 2 securities (assets). Let St :¼ ðSt1 ; . . . ; StK Þ denote the vector of security prices at time t
0; 1; 2; . . . : Assume that Stk > 0 for each t and k, and define by Rtk :¼

Stk k St1

ðk ¼ 1; 2; . . . ; K ; t ¼ 1; 2; . . .Þ

ð4:1Þ

the (gross) return on asset k over the time period (t1, t]. Let Rt :¼ ðRt1 ; . . . ; RtK Þ. At each time period t, an investor chooses a portfolio ht ¼ ðht1 ; . . . ; htK Þ, where htk is the number of units of asset k in the portfolio ht. Generally, ht might depend on the observed values of the price vectors S0 ; S1 ; . . . ; St . A sequence H ¼ ðh0 ; h1 ; . . .Þ specifying a portfolio ht ¼ ht ðS0 ; . . . ; St Þ at each time t as a measurable function of S0 ; S1 ; . . . ; St is called a trading strategy. If not otherwise stated, we will consider only those trading strategies for which htk  0, thus excluding short sales of assets (htk can take on all non-negative real values). One can specify trading strategies in terms of investment proportions (or portfolio weights). Suppose that for each t ¼ 1; 2; . . . ; we are given a vector lt ¼ ðlt1 ; . . . ; ltK Þ in the unit simplex ( D :¼

) K X  1  k K k l ¼1 : l ¼ l ; . . . ; l : l  0; k¼1

The vector lt is assumed to be a measurable function of S0 ; . . . ; St . Given an initial portfolio h0 (specified by a non-negative non-zero vector), we can construct a trading strategy H recursively by the formula htk ¼ ltk St ht1 =Stk

ðk ¼ 1; 2; . . . ; K ; t ¼ 1; 2; . . .Þ:

ð4:2Þ

PK k expresses the value of the portfolio ht1 in Here the scalar product St ht1 ¼ k¼1 Stk ht1 k terms of the prices St at time t. An investor following the strategy (4.2) rebalances (without transaction costs) the portfolio ht1 at time t so that the available wealth St ht1 is

70 j CHAPTER 4

distributed across the assets k
1, 2, . . . , K according to the proportions lt1 ; . . . ; ltK . It is immediate from (4.2) that St ht ¼ St ht1 ;

t ¼ 1; 2; . . . ;

ð4:3Þ

i.e. the strategy H is self-financing. If a strategy is self-financing, then the relations (4.2) and Stk htk ¼ ltk St ht ;

t ¼ 1; 2; . . . ;

ð4:4Þ

are equivalent. If the vectors of proportions lt are fixed (do not depend on time and on the price process), i.e. lt ¼ l ¼ ðl 1 ; . . . ; l K Þ 2 D, then the strategy H defined recursively by htk ¼ l k St ht1 =Stk

ðk ¼ 1; 2; . . . ; K ; t ¼ 1; 2; . . .Þ

ð4:5Þ

is called a constant proportions strategy (or a fixed-mix strategy) with vector of proportions l ¼ ðl 1 ; . . . ; l K Þ. If l k > 0 for each k, then H is said to be completely mixed. We will assume that the price vectors St , and hence the return vectors Rt , are random, i.e. they change in time as stochastic processes. Then the trading strategy ht ; t ¼ 0; 1; 2; . . . , generated by the investment rule (4.2) and the value Vt ¼ St ht ; t ¼ 0; 1; 2; . . . , of the portfolio ht are stochastic processes as well. We are interested in the asymptotic behaviour of Vt as t 0 for constant proportions strategies. We will assume: (R) The vector stochastic process Rt ; t ¼ 1; 2; . . . ; is stationary and ergodic. The expected values Ej ln Rtk j, k
1,2, . . . , K, are finite. Recall that a stochastic process R1 ; R2 ; . . . is called stationary if, for any m ¼ 0; 1; 2; . . . and any measurable function fðx0 ; x1 ; . . . ; xm Þ, the distribution of the random variable ft :¼ fðRt ; Rtþ1 ; . . . ; Rtþm Þ (t
1, 2, . . .) does not depend on t. According to this definition, all probabilistic characteristics of the process Rt are time-invariant. If Rt is stationary, then for any measurable function f for which EjfðRt ; Rtþ1 ; . . . ; Rtþm Þj < 1, the averages f1 þ þ ft t

ð4:6Þ

converge almost surely (a.s.) as t 0 (Birkhoff ’s ergodic theorem—see, e.g. Billingsley 1965). If the limit of all averages of the form (4.6) is non-random (equal to a constant a.s.), then the process Rt is called ergodic. In this case, the above limit is equal a.s. to the expectation Eft , which does not depend on t by virtue of stationarity of Rt. An example of a stationary ergodic process is a sequence of independent identically distributed (i.i.d.) random variables. To avoid misunderstandings, we emphasize that

VOLATILITY-INDUCED FINANCIAL GROWTH

j

71

Brownian motion and a random walk are not stationary. According to the conventional probabilistic terminology, these Markov processes are (time) homogeneous. We have Stk ¼ S0k R1k . . . Rtk , where (according to (R)) the random sequence Rtk is stationary. This assumption on the structure of the price process is a fundamental hypothesis commonly accepted in finance. Moreover, it is quite often assumed that the random variables Rtk ; t ¼ 1; 2; . . . are independent, i.e. the price process Stk forms a geometric random walk. This postulate, which is much stronger than the hypothesis of stationarity of Rtk , lies at the heart of the classical theory of asset pricing (Black, Scholes, Merton), see e.g. Luenberger (1998). By virtue of Birkhoff ’s ergodic theorem, we have

lim

t!1

t 1 1X ln Stk ¼ lim ln Rjk ¼ E ln Rtk ða:s:Þ t!1 t t j¼1

ð4:7Þ

for each k
1, 2, . . ., K. This means that the price of each asset k has almost surely a welldefined and finite (asymptotic, exponential) growth rate, which turns out to be equal a.s. to the expectation r k :¼ E ln Rtk , the drift of this asset’s price. The drift can be positive, zero or negative. It does not depend on t in view of the stationarity of Rt. Let H ¼ ðh0 ; h1 ; . . .Þ be a trading strategy. If the limit lim

t!1

1 lnðSt ht Þ t

exists, it is called the (asymptotic, exponential) growth rate of the strategy H. We now formulate central results of this paper—Theorems 4.1 and 4.2. In these theorems, H ¼ ðh0 ; h1 ; . . .Þ is a constant proportions strategy with some vector of proportions l ¼ ðl 1 ; . . . ; l K Þ 2 D and a non-zero initial portfolio h0 ]0. In Theorems 4.1 and 4.2, we assume that the following condition holds: (V)

With strictly positive probability, Stk Stm

6¼

k St1 m St1

for some 1 k; m K and t  1:

Theorem 4.1: If all the coordinates lk of the vector l are strictly positive, i.e. the strategy H is completely mixed, then the growth rate of the strategy H is almost surely equal to a constant P which is strictly greater than k l k r k , where rk is the drift of asset k. Condition (V) is a very mild assumption of volatility of the price process. This condition does not hold if and only if, with probability one, the ratio Stk =Stm of the prices of any two assets k and m does not depend on t. Thus condition (V) fails to hold if and only if the relative prices of the assets are constant in time (a.s.).

72 j CHAPTER 4

We are primarily interested in the situation when all the assets under consideration have the same drift and hence a.s. the same asymptotic growth rate: (R1) There exists a number r such that, for each k ¼ 1; . . . ; K , we have E ln Rtk ¼ r. From Theorem 4.1, we immediately obtain the following result. Theorem 4.2: Under assumption (R1), the growth rate of the strategy H is almost surely strictly greater than the growth rate of each individual asset. In the context of Theorem 4.2, the volatility of the price process appears to be the only cause for any completely mixed constant proportions strategy to grow at a rate strictly greater than r, the growth rate of each particular asset. This result contradicts conventional finance theory, where the volatility of asset prices is usually regarded as an impediment to financial growth. The result shows that in the present context volatility serves as an endogenous source of its acceleration.

4.3

PROOFS OF THE MAIN RESULTS AND THEIR EXPLANATION

We first observe that if a strategy H is generated according to formula (4.2) by a sequence l1 ; l2 ; . . . of vectors of investment proportions, then Vt ¼ St ht ¼

K X

m Stm ht1

m¼1

¼

K X

Stm

K X Stm m m ¼ St1 ht1 m m¼1 St1

m lt1 St1 ht1 m m¼1 St1

¼ Vt1

K X

m Rtm lt1

m¼1

¼ ðRt lt1 ÞVt1

ð4:8Þ

for each t ] 2, and so Vt ¼ ðRt lt1 ÞðRt1 lt2 Þ . . . ðR2 l1 ÞV1 ; t  2:

ð4:9Þ

Proof of Theorem 4.1: By virtue of (4.9), we have   Vt ¼ V1 ðR1 lÞ1 ðR1 lÞðR2 lÞ . . . ðRt lÞ;

ð4:10Þ

and so

lim

t!1

1 t

ln Vt ¼ lim

t!1

t 1X

t

j¼1

lnðRj lÞ ¼ E lnðRt lÞ ða:s:Þ

ð4:11Þ

VOLATILITY-INDUCED FINANCIAL GROWTH

j

73

by virtue of Birkhoff ’s ergodic theorem. It remains to show that if assumption (V) holds, PK then E lnðRt lÞ > k¼1 l k r k . To this end observe that condition (V) is equivalent to the following one. (V1) For some t ] 1 (and hence, by virtue of stationarity, for each t ] 1), the probability P Rtk 6¼ Rtm

for some 1 k; m K



is strictly positive. k m k m =St1 if and only if Stk =St1 6¼ Stm =St1 , which can be Indeed, we have Stk =Stm 6¼ St1 k m written as Rt 6¼ Rt . Denote by dt the random variable that is equal to 1 if the event fRtk 6¼ Rtm for some 1 k; m K g occurs and 0 otherwise. Condition (V) means that Pfmaxt1 dt ¼ 1g > 0, while (V1) states that, for some t (and hence for each t), Pfdt ¼ 1g > 0. The latter property is equivalent to the former because

 [ 1 max dt ¼ 1 ¼ fdt ¼ 1g: t1

t¼1

By using Jensen’s inequality and (V1), we find that ln

K X

Rtk l k >

k¼1

K X

  l k ln Rtk

k¼1

with strictly positive probability, while the non-strict inequality holds always. Consequently, E lnðRt lÞ >

K X k¼1

K   X l k E ln Rtk ¼ l kr k;

ð4:12Þ

k¼1

which completes the proof. I The above considerations yield a rigorous proof of the fact of volatility induced growth. But what is the intuition, the underlying fundamental reason for it? We have only one explanation, which is nothing but a repetition in one phrase of the idea of the above proof. If Rt1 ; . . . ; RtK are random returns of assets k
1, 2, . . . , K, then the asymptotic growth rates of these assets are E ln Rtk , while the asymptotic growth rate of a constant P k P l E lnðRtk Þ by proportions strategy is E lnð l k Rtk Þ, which is strictly greater than Jensen’s inequality—because the function ln x is strictly concave. It would be nice, however, to give a general common-sense explanation of volatilityinduced growth, without using such terms as a ‘strictly convex function,’ ‘Jensen’s inequality,’ etc. One can, indeed, find in the literature explanations of examples of volatility pumping based on the following reasoning (see e.g. Fernholz and Shay 1982;

74 j CHAPTER 4

Luenberger 1998). The reason for growth lies allegedly in the fact that constant proportions always force one to ‘buy low and sell high’—the common sense dictum of all trading. Those assets whose prices have risen from the last rebalance date will be overweighted in the portfolio, and their holdings must be reduced to meet the required proportions and to be replaced in part by assets whose prices have fallen and whose holdings must therefore be increased. Obviously, for this mechanism to work the prices must change in time; if they are constant, one cannot get any profit from trading. We have, alas, repeated this reasoning ourselves (e.g. in Evstigneev and Schenk-Hoppe´ 2002 and in an earlier version of the present chapter), but somewhat deeper reflection on this issue inevitably leads to the conclusion that the above argument does not explain everything and raises more questions than it gives answers. For example, what is the meaning of ‘high’ and ‘low?’ If the price follows a geometric random walk, the set of its values is generally unbounded, and for every value there is a larger value. One can say that ‘high’ and ‘low’ should be understood in relative terms, based on the comparison of the prices today and yesterday. Fine, but what if the prices of all the assets increase or decrease simultaneously? Thus, the above argument, to be made valid, should be at least relativized, both with respect to time and the assets. However, a more substantial lacuna in such reasoning is that it does not reflect the assumption of constancy of investment proportions. This leads to the question: what will happen if the ‘common sense dictum of all trading’ is pushed to the extreme and the portfolio is rebalanced so as to sell all those assets that gain value and buy only those ones which lose it? Assume, for example, that there are two assets, the price St1 of the first (riskless) is always 1, and the price St2 of the second (risky) follows a geometric random walk, so that the gross return on it can be either 2 or 1/2 with equal probabilities. Suppose the investor sells the second asset and invests all wealth in the first if the price St2 goes up and performs the converse operation, betting all wealth on the risky asset, if St2 goes down. Then the sequence lt ¼ ðlt1 ; lt2 Þ of the vectors of investment proportions will be i.i.d. with values (0, 1) and (1, 0) taken on with equal probabilities. Furthermore, lt1 will be independent of Rt. By virtue of (4.9), the growth rate of the portfolio value for this strategy is equal to E lnðRt lt1 Þ ¼ ½lnð0 1 þ 1 2Þ þ lnð0 1 þ 1 12Þ þ lnð1 1 þ 0 2Þþ lnð1 1 þ 0 12Þ=4 ¼ 0, which is the same as the growth rate of each of the two assets k
1, 2 and is strictly less than the growth rate of any completely mixed constant proportions strategy.

4.4

STATIONARY MARKETS: PUZZLES AND MISCONCEPTIONS

Consider a market where the price process St (and not only the process of asset returns Rt) is ergodic and stationary and where Ej ln Stk j < 1. This situation is a special case of stationary returns, because if the vector process St is stationary, then the process Rt is stationary as well. In this case the growth rate of each asset is zero, k ¼ 0; E ln Rtk ¼ E ln Stk  E ln St1

VOLATILITY-INDUCED FINANCIAL GROWTH

j

75

while, as we have seen, any completely mixed constant proportions strategy grows at a strictly positive exponential rate. The assumption of stationarity of asset prices, perhaps after some detrending, seems plausible when modelling currency markets (Kabanov 1999; Dempster et al. 2003). Then the ‘prices’ are determined by exchange rates of all the currencies with respect to some selected reference currency. We performed a casual experiment, asking numerous colleagues (in private, at seminars and at conferences) to promptly guess the correct answer to the following question. Question 4.1: Suppose vectors of asset prices St ¼ (St1 ; . . . ; StK ) fluctuate randomly, forming a stationary stochastic process (assume even that St are i.i.d.). Consider a fixedmix self-financing investment strategy prescribing rebalancing one’s portfolio at each of the dates t
1, 2, . . . , so as to keep equal investment proportions of wealth in all the assets. What will happen with the portfolio value in the long run, as t0? What will be its tendency: (a) to decrease; (b) to increase; or (c) to fluctuate randomly, converging in one sense or another to a stationary process? The audience of our respondents was quite broad and professional, but practically nobody succeeded in guessing the correct answer, (b). Among those who expressed a clear guess, nearly all selected (c). There were also a couple of respondents who decided to bet on (a). Common intuition suggests that if the market is stationary, then the portfolio value Vt for a constant proportions strategy must converge in one sense or another to a stationary process. The usual intuitive argument in support of this conjecture appeals to the selffinancing property (3). The self-financing constraint seems to exclude possibilities of unbounded growth. This argument is also substantiated by the fact that in the deterministic case both the prices and the portfolio value are constant. This way of reasoning makes the answer (c) to the above question more plausible a priori than the others. It might seem surprising that the wrong guess (c) has been put forward even by those who have known about examples of volatility pumping for a long time. The reason for this might lie in the non-traditional character of the setting where not only the asset returns but the prices themselves are stationary. Moreover, the phenomenon of volatility-induced growth is more paradoxical in the case of stationary prices, where growth emerges ‘from nothing.’ In the conventional setting of stationary returns, volatility serves as the cause of an acceleration of growth, rather than its emergence from prices with zero growth rates. A typical way of understanding the correct answer to Question 4.1 is to reduce it to something well known that is apparently relevant. A good candidate for this is the concept of arbitrage. Getting something from nothing as a result of an arbitrage opportunity seems to be similar to the emergence of growth in a stationary setting where there are no obvious sources for growth. As long as we deal with an infinite time horizon, we would have to consider some kind of asymptotic arbitrage (e.g. Ross 1976; Huberman 1982; Kabanov and Kramkov 1994; Klein and Schachermayer 1996). However, all known concepts of this kind are much weaker than what we would need in the present context. According to our results, growth is exponentially fast, unbounded wealth is achieved with probability one, and the effect of

76 j CHAPTER 4

growth is demonstrated for specific (constant proportions) strategies. None of these properties can be directly deduced from asymptotic arbitrage. Thus there are no convincing arguments showing that volatility-induced growth in stationary markets can be derived from, or explained by, asymptotic arbitrage over an infinite time horizon. But what can be said about relations between stationarity and arbitrage over finite time intervals? As is known, there are no arbitrage opportunities (over a finite time horizon) if and only if there exists an equivalent martingale measure. A stationary process can be viewed as an ‘antipodal concept’ to the notion of a martingale. This might lead to the conjecture that in a stationary market arbitrage is a typical situation. Is this true or not? Formally, the question can be stated as follows. Question 4.2: Suppose the process St ¼ ðSt1 ; . . . ; StK Þ of the vectors of asset prices is stationary, and moreover, assume that the vectors St are i.i.d. Furthermore, suppose the first asset k
1 is riskless and its price St1 is equal to one. Does this market have arbitrage opportunities over a finite time horizon? When asking this question, we assume that the market is frictionless and that there are no portfolio constraints. In particular, all short sales are allowed. An arbitrage opportunity over a time horizon t ¼ 0; . . . ; T is understood in the conventional sense. It means the existence of a self-financing trading strategy ðh0 ; . . . ; hT Þ such that S0 h0 ¼ 0, ST hT  0 a.s. and PfST hT > 0g > 0. The answer to this question, as well as to the previous one, is practically never guessed immediately. Surprisingly, the answer depends, roughly speaking, on whether the distribution of the random vector S~t :¼ ðSt2 ; . . . ; StK Þ of prices of the risky assets is continuous or discrete. For example, if S~t takes on a finite number of values, then an arbitrage opportunity exists. If the distribution of S~t is continuous, there are no arbitrage opportunities. More precisely, the result is as follows. Let G be the support of the distribution of the random vector S~t (which is assumed to be non-degenerate) and let F:
cl co G be the closure of the convex hull of G. Denote by 1rF the relative boundary of F, i.e. the boundary of the convex set F in the smallest linear manifold containing F. Theorem 4.3: If PfS~t 2 @r F g ¼ 0, then for any T there are no arbitrage opportunities over the time horizon of length T. If PfS~t 2 @r Fg > 0, then for each T there is an arbitrage opportunity over the time horizon of length T. For a proof of this result see Evstigneev and Kapoor (2006).

4.5

GROWTH ACCELERATION, VOLATILITY REDUCTION AND NOISE-INDUCED STABILITY

The questions we analyse in this section stem from an example of volatility pumping considered originally by Fernholz and Shay (1982) and later others (e.g. Luenberger 1998). The framework for this example is the well-known continuous-time model developed by Merton and others, in which the price processes Stk ðt  0Þ of two assets k
1, 2 are supposed to be solutions to the stochastic differential equations dStk =Stk ¼ mk dt þ sk dWt k , where the Wt k are independent (standard) Wiener processes and S0k ¼ 1. As is well known,

VOLATILITY-INDUCED FINANCIAL GROWTH

j

77

these equations admit explicit solutions Stk ¼ exp½mk t  ðsk2 =2Þt þ sk Wt k . Given some y 2 ð0; 1Þ, the value Vt of the constant proportions portfolio prescribing investing the proportions u and 1  y of wealth into assets k
1, 2 is the solution to the equation dVt Vt

¼ ½ym1 þ ð1  yÞm2 dt þ ys1 dWt 1 þ ð1  yÞs2 dWt 2 :

Equivalently, Vt can be represented as the solution to the equation dt þ s dWt , where m  :¼ ym1 þ ð1  yÞm2 , s  2 :¼ ðys1 Þ 2 þ ½ð1  yÞs2  2 and dVt =Vt ¼ m Wt , and so the growth mt  ð s 2 =2Þt þ s Wt is a standard Wiener process. Thus, Vt ¼ exp½   ð . In s 2 =2Þ and s rate and the volatility of the portfolio value process Vt are given by m particular, if m1 ¼ m2 ¼ m and s1 ¼ s2 ¼ s, then the growth rate and the volatility of Vt are equal to 2

m  ð s =2Þ

and

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  ¼ s y 2 þ ð1  yÞ 2 < s; s

ð4:13Þ

while for each individual asset the growth rate and the volatility are m  ðs 2 =2Þ and s, respectively. Thus, in this example, the use of a constant proportions strategy prescribing investing in a mixture of two assets leads (due to diversification) to an increase of the growth rate and to a simultaneous decrease of the volatility. When looking at the expressions in (4.13), the temptation arises even to say that the volatility reduction is the cause of volatilityinduced growth. Indeed, the growth rate m  ð s 2 =2Þ is greater than the growth rate 2  < s. This suggests speculation along the following lines. Volatility m  ðs =2Þ because s is something like energy. When constructing a mixed portfolio, it converts into growth and therefore itself decreases. The greater the volatility reduction, the higher the growth acceleration. Do such speculations have any grounds in the general case, or do they have a justification only in the above example? To formalize this question and answer it, let us return to the discrete time-framework we deal with in this paper. Suppose there are two assets with i.i.d. vectors of returns Rt ¼ ðRt1 ; Rt2 Þ. Let ðx; ZÞ :¼ ðR11 ; R12 Þ and assume, to avoid technicalities, that the random vector ðx; ZÞ takes on a finite number of values and is strictly positive. The value Vt of the portfolio generated by a fixed-mix strategy with proportions x and 1 x (0 Bx B1) is computed according to the formula V t ¼ V1

t h Y

i xRj1 þ ð1  xÞRj2 ;

t  2;

j¼2

see (4.9). The growth rate of this process and its volatilityffi are given, respectively, by the pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi expectation E ln zx and the standard deviation Var ln zx of the random variable ln zx , where zx :¼ ln½xx þ ð1  xÞZ. We know from the above analysis that the growth rate increases when mixing assets with the same growth rate. What can be said about volatility? Specifically, let us consider the following question.

78 j CHAPTER 4

Question 4.3: (a) Suppose Var ln x ¼ Var ln Z. Is it true that Var ln½xx þ ð1  xÞZ

Var ln x when x  (0,1)? (b) More generally, is it true that Var ln½xx þ ð1  xÞZ maxðVar ln x; Var ln ZÞ for x  (0,1)? Query (b) asks whether the logarithmic variance is a quasi-convex functional. Questions (a) and (b) can also be stated for volatility defined as the square root of logarithmic variance. They will have the same answers because the square root is a strictly monotone function. Positive answers to these questions would substantiate the above conjecture of volatility reduction—negative, refute it. It turns out that in general (without additional assumptions on j and h) the above questions 4.3(a) and 4.3(b) have negative answers. To show this consider two i.i.d. random variables U and V with values 1 and a  0 realized with equal probabilities. Consider the function f ðyÞ :¼ Var ln½yU þ ð1  yÞV ;

y 2 ½0; 1:

ð4:14Þ

By evaluating the first and the second derivatives of this function at y
1/2, one can show the following. There exist some numbers 0 < a < 1 and aþ > 1 such that the function f(y) attains its minimum at the point y
1/2 when a belongs to the closed interval ½a ; aþ  and it has a local maximum (!) at y
1/2 when a does not belong to this interval. The numbers a  and a are given by a ¼ 2e 4  1 

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 ð2e 4  1Þ  1;

where a  :0.0046 and a :216.388. If a 2 ½a ; aþ , the function f (y) is convex, but if a2 = ½a ; aþ , its graph has the shape illustrated in Figure 4.1. Fix any a for which the graph of f (y) looks like the one depicted in Figure 4.1. Consider any number y0 B1/2 which is greater than the smallest local minimum of f (y) and define x :¼ y0 U þ ð1  y0 ÞV and Z :¼ y0 V þ ð1  y0 ÞU . (U and V may be interpreted as ‘factors’ on which the returns j and h on the two assets depend.) Then Var ln½ðx þ ZÞ=2 > Var ln x ¼ Var ln Z, which yields a negative answer both to (a) and (b). In this f (y)

y 0

1/2

FIGURE 4.1 Graph of the function f (y) in Equation (4.14) for a
104.

1

VOLATILITY-INDUCED FINANCIAL GROWTH

j

79

example j and h are dependent. It would be of interest to investigate questions (a) and (b) for general independent random variables j and h. It can be shown that the answer to (b) is positive if one of the variables j and h is constant. But even in this case the function Var ln½xx þ ð1  xÞZ is not necessarily convex: it may have an inflection point in (0, 1), which can be easily shown by examples involving two-valued random variables. Thus it can happen that a mixed portfolio may have a greater volatility than each of the assets from which it has been constructed. Consequently, the above conjecture and the ‘energy interpretation’ of volatility are generally not valid. It is interesting, however, to find additional conditions under which assertions regarding volatility reduction hold true. In this connection, we can assert the following fact. Theorem 4.4: Let U and V be independent random variables bounded above and below by strictly positive constants. If U is not constant, then Var ln½ yU þ ð1  yÞV  < Var In U for all y  (0,1) sufficiently close to 1. Volatility can be regarded as a quantitative measure of instability of the portfolio value. The above result shows that small independent noise can reduce volatility. This result is akin to a number of known facts about noise-induced stability, e.g. Abbott (2001) and Mielke (2000). An analysis of links between the topic of the present work and results about stability under random noise might constitute an interesting theme for further research. Proof of Theorem 4.4: This can be obtained by evaluating the derivative of the function f(y) defined by (4.14) at y
1. Basic computations show that f 0 ð1Þ ¼ 2ðEV ÞðEe Z Z þ Ee Z EZÞ;

ð4:15Þ

where Z
ln U. The assertion of the theorem is valid because f 0 ð1Þ > 0. The verification of this inequality is based on the following fact. If f(z) is a function on ( ,) with f 0 ðzÞ > 0, then E ½ZfðZÞ > ðEZÞEfðZÞ

ð4:16Þ

for any non-constant bounded random variable Z. This fact R x follows from Jensen’s inequality applied to the strictly convex function cðxÞ :¼ 0 fðzÞdz and from the relation cðyÞ  cðxÞ  fðxÞðy  xÞ (to obtain (16) put x:
Z, y:
EZ and take the expectation). By applying (16) to fðzÞ :¼ e z , we find that the expression in (4.15) is positive. I

4.6

VOLATILITY-INDUCED GROWTH UNDER SMALL TRANSACTION COSTS

We now assume that, in the market under consideration (see Section 4.2), there are transaction costs for buying and selling assets. When selling x units of asset k at time t, one receives the amount ð1  ek ÞStk x (rather than Stk x as in the frictionless case). To buy x units of asset k, one has to pay ð1 þ eþk ÞStk x. The numbers ek ; eþk  0, k
1, 2, . . . , K

80 j CHAPTER 4

(the transaction cost rates) are assumed given. In this market, a trading strategy H ¼ ðh0 ; h1 ; . . .Þ is self-financing if K X 

1þ

eþk

K X  k k     k  k St ht  ht1 þ 1  ek Stk ht1  htk þ ;

k¼1

t  1;

ð4:17Þ

k¼1

where xþ ¼ maxfx; 0g. Inequality (4.17) means that asset purchases can be made only at the expense of asset sales. Relation (4.17) is equivalent to K X

Stk



htk



k ht1





k¼1

K X

eþk Stk

k¼1



htk





k ht1  þ

K X

 k  ek Stk ht1  htk þ

k¼1

(which follows from the identity xþ  ðxÞþ ¼ x). Therefore, if the portfolio ht differs from the portfolio ht1 in at least one position k for which eþk 6¼ 0 and ek 6¼ 0, then St ht < St ht1 . Thus, under transaction costs, portfolio rebalancing typically leads to a loss of wealth. The number St ht =St ht1 1 is called the loss coefficient (of the portfolio strategy H at time t). We say that H ¼ ðh0 ; h1 ; . . .Þ is a constant proportions strategy with vector of proportions l ¼ ðl 1 ; . . . ; l K Þ 2 D if Stk htk ¼ l k St ht for all k ¼ 1; 2; . . . ; K and t
1, 2, . . . (cf. (4)). Let d  (0, 1) be a constant. Given a vector of proportions l ¼ ðl 1 ; . . . ; l K Þ 2 D and a non-zero initial portfolio h0 ]0, define recursively htk ¼

ð1  dÞl k St ht1 Stk

ðk ¼ 1; 2; . . . ; K ; t ¼ 1; 2; . . .Þ:

ð4:18Þ

This rule defines a trading strategy with constant investment proportions l 1 ; . . . ; l K and a constant loss coefficient 1d. We will call it the (d, l)-strategy. Theorem 4.5 below extends the results of Theorems 4.1 and 4.2 to the model with transaction costs. As before, we assume that hypotheses (R) and (V) hold. Theorem 4.5: Let l ¼ ðl 1 ; . . . ; l K Þ 2 D be a strictly positive vector. If d  (0,1) is small enough, then the (d, l)-strategy H defined by (4.18) has a growth rate strictly greater than PK k k 1 K k¼1 l r (a.s.), and so if r ¼ ¼ r ¼ r, then the growth rate of H is strictly greater than r (a.s.). Further, if the transaction cost rates ek ; eþk  0, k
1, 2, . . ., K, are small enough (in particular, if they do not exceed du2), then the strategy H is self-financing. The purpose of this theorem is to demonstrate that the results on volatility-induced growth remain valid under small transaction costs. In contrast with a number of the questions we considered previously, the answer to the question we pose here is quite predictable and does not contradict intuition. We deal in Theorem 4.5 with constant proportions strategies of a special form—those for which the loss rate is constant (and small enough). We are again not concerned with the question of optimality of such strategies for one criterion or another.

VOLATILITY-INDUCED FINANCIAL GROWTH

j

81

Proof of Theorem 4.5: We first observe that the growth rate of the strategy H is equal to E ln½ð1  dÞRt l. This fact is proved exactly like (4.11) (simply multiply the vectors of proportions in (4.8), (4.9), (4.10) and (4.11) by (1d)). According to (4.12), we have PK E lnðRt lÞ > k¼1 l k r k . This inequality will remain valid if l is replaced by (1d)l, provided d  (0, 1) is small enough. Fix any such d  (0, 1). Denote by o the greatest of the numbers ek ; eþk . It remains to show that H is self-financing when e d=2. To this end we note that inequality (4.17) is implied by K K X X  k   k  k k ð1 þ eÞSt ht  ht1 þ ð1  eÞStk ht1  htk þ ; k¼1

k¼1

which is equivalent to K X  k k  S h  S k h k  S ðh  h Þ: e t t1 t t t t t1

ð4:19Þ

k¼1

In view of (4.18), the right-hand side of the last inequality is equal to dSt ht1, and the lefthand side can be estimated as follows: e

K X k¼1

k jð1  dÞlk St ht1  Stk ht1 j e

K X

ð1  dÞlk St ht1 þ e

k¼1

K X

k Stk ht1

k¼1

¼ eð1  dÞSt ht1 þ eSt ht1 2eSt ht1 : Consequently, if 0 e d=2, then the strategy H is self-financing.

4.7

I

GROWTH UNDER TRANSACTION COSTS: AN EXAMPLE

In this section we consider an example (a binomial model) in which quantitative estimates for the size of the transaction costs needed for the validity of the result on volatilityinduced growth can be provided. Suppose that there are two assets k
1, 2: one riskless and one risky. The price of the former is constant and equal to 1. The price of the latter follows a geometric random walk. It can either jump up by u  1 or down by u1 with equal probabilities. Thus both securities have zero growth rates. Suppose the investor pursues the constant proportions strategy prescribing to keep 50) of wealth in each of the securities. There are no transaction costs for buying and selling the riskless asset, but there is a transaction cost rate for buying and selling the risky asset of e 2 ð0; 1Þ. Assume the investor’s portfolio at time t  1 contains v units of cash; then the value of the risky position of the portfolio must be also equal to v. At time t, the riskless position of the portfolio will remain the same, and the value of the risky position will become either uv or u1v with equal probability. In the former case, the investor rebalances his/her portfolio by selling an amount of the risky asset worth A so that v þ ð1  eÞA ¼ vu  A:

ð4:20Þ

82 j CHAPTER 4

by selling an amount of the risky asset of value A in the current prices, the investor receives ð1  eÞA, and this sum of cash is added to the riskless position of the portfolio. After rebalancing, the values of both portfolio positions must be equal, which is expressed in 1 (4.20). From (4.20) we obtain A ¼ vðu  1Þð2  eÞ . The positions of the new (rebalanced) portfolio, measured in terms of their current values, are equal to 1 v þ ð1  eÞA= v[1 ð1  eÞð2  eÞ ðu  1Þ. In the latter case (when the value of the risky position becomes u 1), the investor buys some amount of the risky asset worth B, for which the amount of cash ð1 þ eÞB is needed, so that v  ð1 þ eÞB ¼ u 1 v þ B: From this, we find B ¼ vðu 1  1Þð2 þ eÞ 1 , and so v  ð1 þ eÞB ¼ v½1 þ ð1 þ eÞ ð2 þ eÞ 1 ðu 1  1Þ. Thus, the portfolio value at each time t is equal to its value at time t  1 multiplied by the random variable j such that Pfx ¼ g 0 g ¼ Pfx ¼ g 00 g ¼ 1=2, where g 0 :¼ 1 þ ð1 þ eÞ ð2 þ eÞ 1 ðu 1  1Þ and g 00 :¼ 1 þ ð1  eÞ  ð2  eÞ 1 ðu  1Þ . Consequently, the asymptotic growth rate of the portfolio value, E ln x ¼ ð1=2Þðln g 0 þ ln g 00 Þ, is equal to ð1=2Þ ln fðe; uÞ, where 

u 1  1 fðe; uÞ :¼ 1 þ ð1 þ eÞ 2þe



 u1 1 þ ð1  eÞ : 2e

We have E ln x > 0, i.e. the phenomenon of volatility induced growth takes place, if fðe; uÞ > 1. For e 2 ½0; 1Þ, this inequality turns out to be equivalent to the following very simple relation: 0 e < ðu  1Þðu þ 1Þ 1 . Thus, given u  1, the asymptotic growth rate of the fixed-mix strategy under consideration is greater than zero if the transaction cost rate o is less than e ðuÞ :¼ ðu  1Þðu þ 1Þ 1 . We call e ðuÞ the threshold transaction cost rate. Volatility-induced growth takes place—in the present example, where the portfolio is rebalanced in every one period2—when 0 e < e ðuÞ. The volatility s of the risky asset under consideration (the standard deviation of its logarithmic return) is equal to ln u. In the above considerations, we assumed that s—or equivalently, u—is fixed, and we examined fðe; uÞ as a function of o. Let us now examine fðe; uÞ as a function of u when the transaction cost rate o is fixed and strictly positive. For the derivative of fðe; uÞ with respect to u, we have fu0 ðe; uÞ

  1þe 1e 2 u : ¼ 4  e2 1 þ e

If u
1, then fu0 ðe; 1Þ < 0. Thus if the volatility of the risky asset is small, the performance of the constant proportions strategy at hand is worse than the performance of each individual asset. This fact refutes the possible conjecture that ‘the higher the volatility, 2

For the optimal timing of rebalancing in markets with transaction costs see, e.g. Aurell and Muratore (2000).

VOLATILITY-INDUCED FINANCIAL GROWTH

j

83

the higher the induced growth rate.’ Further, the derivative fu0 ðe; uÞ is negative when 1=2 1=2 ð1 þ eÞ . For u ¼ uðeÞ the asymptotic growth u 2 ½0; uðeÞÞ, where uðeÞ :¼ ð1  eÞ rate of the constant proportions strategy at hand attains its minimum value. For those values of u which are greater than uðeÞ, the growth rate increases, and it tends to infinity as u 0. Thus, although the assertion ‘the greater the volatility, the greater the induced growth rate’ is not valid always, it is valid (in the present example) under the additional assumption that the volatility is large enough.

4.8

CONCLUSION

In this chapter we have established the surprising result that when asset returns are stationary ergodic, their volatility, together with any fixed-mix trading strategy, generates a portfolio growth rate in excess of the individual asset growth rates. As a consequence, even if the growth rates of the individual securities all have mean zero, the value of a fixedmix portfolio tends to infinity with probability one. By contrast with the 25 years in which the effects of ‘volatility pumping’ have been investigated in the literature by example, our results are quite general. They are obtained under assumptions which accommodate virtually all the empirical market return properties discussed in the literature. We have in this chapter also dispelled the notion that the demonstrated acceleration of portfolio growth is simply a matter of ‘buying lower and selling higher.’ The example of Section 4.3 shows that our result depends critically on rebalancing to an arbitrary fixed mix of portfolio proportions. Any such mix defines the relative magnitudes of individual asset returns realized from volatility effects. This observation and our analysis of links between growth, arbitrage and noise-induced stability suggest that financial growth driven by volatility is a subtle and delicate phenomenon.

ACKNOWLEDGEMENTS We are indebted to numerous colleagues who had enough patience to participate in our ‘intuition tests,’ the results of which are discussed in Section 4.4 of this chapter. Our comparative analysis of volatility-induced growth and arbitrage (also contained in Section 4.4) was motivated by Walter Schachermayer’s comments at a conference on Mathematical Finance in Paris in 2003. Albert N. Shiryaev deserves our thanks for a helpful discussion of some mathematical questions related to this work. Financial support by the Swiss National Centre of Competence in Research ‘Financial Valuation and Risk Management’ (NCCR FINRISK) is gratefully acknowledged.

REFERENCES Abbott, D., Overview: Unsolved problems of noise and fluctuations. Chaos, 2001, 11, 526/538. Algoet, P.H. and Cover, T.M., Asymptotic optimality and asymptotic equipartition properties of logoptimum investment. Ann. Probab., 1988, 16, 876/898. Aurell, E., Baviera, R., Hammarlid, O., Serva, M. and Vulpiani, A., A general methodology to price and hedge derivatives in incomplete markets. Int. J. Theoret. Appl. Finance, 2000, 3, 1 /24. Aurell, E. and Muratore, P., Financial friction and multiplicative Markov market games. Int. J. Theoret. Appl. Finance, 2000, 3, 501/510.

84 j CHAPTER 4

Billingsley, P., Ergodic Theory and Information, 1965 (Wiley: New York). Breiman, L., Optimal gambling systems for favorable games, in Fourth Berkeley Symposium on Mathematical Statistics and Probability, Vol. 1, 1961, pp. 65 /78. Cover, T.M., Universal portfolios. Math. Finance, 1991, 1, 1 /29. Dempster, M.A.H., Evstigneev, I.V. and Schenk-Hoppe´ , K.R., Exponential growth of fixed-mix strategies in stationary asset markets. Finance Stochast., 2003, 7, 263 /276. Dries, D., Ilhan, A., Mulvey, J., Simsek, K.D. and Sircar, R., Trend-following hedge funds and multiperiod asset allocation. Quant. Finance, 2002, 2, 354/361. Evstigneev, I.V. and Kapoor, D., Arbitrage in stationary markets. Discussion Paper 0608, 2006, (School of Economic Studies, University of Manchester). Evstigneev, I.V. and Schenk-Hoppe´ , K.R., From rags to riches: On constant proportions investment strategies. Int. J. Theoret. Appl. Finance, 2002, 5, 563/573. Fernholz, R. and Shay, B., Stochastic portfolio theory and stock market equilibrium. J. Finance, 1982, 37, 615/624. Fo¨ llmer, H. and Schied, A., Stochastic Finance—An Introduction in Discrete Time, 2002 (Walter de Gruyter: Berlin). Hakansson, N.H. and Ziemba, W.T., Capital growth theory, in Handbooks in Operations Research and Management Science, Volume 9, Finance, edited by R.A. Jarrow, V. Maksimovic and W.T. Ziemba, pp. 65/86, 1995 (Elsevier: Amsterdam). Harmer, G.P. and Abbott, D., Parrondo’s paradox. Stat. Sci., 1999, 14, 206/213. Huberman, G., A simple approach to arbitrage pricing theory. J. Econ. Theory, 1982, 28, 183/191. Kabanov, YuM., Hedging and liquidation under transaction costs in currency markets. Finance Stochast., 1999, 3, 237 /248. Kabanov, YuM. and Kramkov, D.A., Large financial markets: Asymptotic arbitrage and contiguity. Theory Probab. Appl., 1994, 39, 222/228. Kelly, J.L., A new interpretation of information rate. Bell Sys. Techn. J., 1956, 35, 917/926. Klein, I. and Schachermayer, W., Asymptotic arbitrage in non-complete large financial markets. Theory Probab. Appl., 1996, 41, 927/334. Li, Y., Growth-security investment strategy for long and short runs. Manag. Sci., 1998, 39, 915 /924. Luenberger, D., Investment Science, 1998 (Oxford University Press: New York). MacLean, L.C., Ziemba, W.T. and Blazenko, G., Growth versus security in dynamic investment analysis. Manag. Sci., 1992, 38, 1562/1585. Mielke, A., Noise induced stability in fluctuating bistable potentials. Phys. Rev. Lett., 2000, 84, 818/821. Mulvey, J.M., Multi-period stochastic optimization models for long-term investors, in Quantitative Analysis in Financial Markets, edited by M. Avellaneda, Vol. 3, pp. 66 /85, 2001 (World Scientific: Singapore). Mulvey, J.M. and Ziemba, W.T., (Eds), Worldwide Asset and Liability Modeling, 1998 (Cambridge University Press: Cambridge, UK). Perold, A.F. and Sharpe, W.F., Dynamic strategies for asset allocation. Financ. Analysts J., 1995, 51, 149/160. Ross, S.A., The arbitrage theory of asset pricing. J. Econ. Theory, 1976, 13, 341/360.

CHAPTER

5

Constant Rebalanced Portfolios and Side-Information E. FAGIUOLI, F. STELLA and A. VENTURA CONTENTS 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Constant Rebalanced Portfolios . . . . . . . . . . 5.3 Online Investment with Side-Information. . . . 5.4 Numerical Experiments . . . . . . . . . . . . . . . . 5.5 Conclusions and Further Research Directions. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5.1

T

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. 85 . 87 . 90 . 99 104 106 106

INTRODUCTION

is one of the basic problems within the research area of computational finance. It has been studied intensively throughout the last 50 years, producing several relevant contributions described in the specialized literature. Portfolio selection originates from the seminal paper of Markowitz (1952), who introduced and motivated the mean-variance investment framework. This conventional approach to portfolio selection, which has received increasing attention, consists of two separate steps. The first step concerns distributional assumptions about the behaviour of stock prices, while the second step is related to the selection of the optimal portfolio depending on some objective function and/or utility function defined according to the investor’s goal. This conceptual model has proved in the past to be useful in spite of the many drawbacks that have been pointed out by finance practitioners, private investors and researchers. Indeed, the first step, related to distributional assumptions concerning the behaviour of stock prices, encounters many difficulties because the future evolution of stock prices is notoriously difficult to predict, while the selection of a distribution class inevitably brings a measure of arbitrariness. These problems become even more evident and dramatic in the case where there are reasons to believe that the process that governs stock price behaviour changes over time. HE PORTFOLIO SELECTION PROBLEM

85

86 j CHAPTER 5

A different approach to portfolio selection, to overcome the main limitations and problems related to the mean variance approach, has been proposed by Cover (1991a, b). The main characteristic of Cover’s approach to portfolio selection is that no distributional assumptions on the sequence of price relatives are required. Indeed, within Cover’s investment framework, portfolio selection is based completely on the sequence of past prices, which is taken as is, with little, if any, statistical processing. No assumptions are made concerning the family of probability distributions that describes the stock prices, or even concerning the existence of such distributions. To emphasize this independence from statistical assumptions, such portfolios are called universal portfolios. It has been shown that such portfolios possess important theoretical properties concerning their asymptotic behaviour and exhibit reasonable finite time behaviour. Indeed, it is well known (Bell and Cover 1980; Algoet and Cover 1988; Cover 1991a) that if the price relatives are independent and identically distributed, the optimal growth rate of wealth is achieved by a constant rebalanced portfolio, i.e. an investment strategy that keeps fixed through time, trading period by trading period, the distribution of wealth among a set of assets. In recent years, constant rebalanced portfolios have received increasing attention (Auer and Warmuth 1995; Herbster and Warmuth 1995; Helmbold et al. 1996; Singer 1997; Browne 1998; Vovk and Watkins 1998; Borodin et al. 2000; Gaivoronski and Stella 2000, 2003) and have also been studied in the case where transaction costs are involved (Blum and Kalai 1998; Evstigneev and Schenk-Hoppe` 2002). It is worth noting that the best constant rebalanced portfolio, as well as the universal portfolio, are designed to deal with the portfolio selection problem in the case where no additional information is available concerning the stock market. However, it is common practice that investors, fund managers and private investors adjust their portfolios, i.e. rebalance, using various sources of information concerning the stock market which can be conveniently summarized by the concept of side-information. A typical example of sideinformation originates from sophisticated trading strategies that often develop signalling algorithms that individuate the nature of the investment opportunity about to be faced. In this context, the side-information is usually considered to be a causal function of past stock market performance. Cover and Ordentlich (1996) presented the state constant rebalanced portfolio, i.e. a sequential investment algorithm that achieves, to first order in the exponent, the same wealth as the best side-information-dependent investment strategy determined in hindsight from observed market and side-information outcomes. The authors, at each trading period t  {1, . . ., n}, used a state constant rebalanced portfolio investment algorithm that invests in the market using one of k distinct portfolios x(1), . . ., x(k) depending on the current state of side-information yt. They established a set of allowable investment actions (sequence of portfolio choices xt), and sought to achieve the same asymptotic growth rate of wealth as the best action in this set, not in any stochastic sense, but uniformly over all possible sequences of price relatives and side-information states. In this chapter we study and analyse the topic proposed by Cover and Ordentlich (1996). Attention is focused on the interplay between constant rebalanced portfolios and side-information. A mathematical framework is proposed for dealing with constant rebalanced portfolios in the case where side-information is available concerning the stock

CONSTANT REBALANCED PORTFOLIOS AND SIDE-INFORMATION

j

87

market. The mathematical framework introduces a new investment strategy, namely the mixture best constant rebalanced portfolio, which directly exploits the available side-information to outperform, in terms of the achieved wealth, the best constant rebalanced portfolio determined in hindsight, i.e. by assuming perfect knowledge of future stock prices. We provide a mathematical comparison of the achieved wealth by means of the best constant rebalanced portfolio and its counterpart, namely the mixture best constant rebalanced portfolio. The mixture best constant rebalanced portfolio is shown to outperform the best constant rebalanced portfolio by an exponential factor in terms of the achieved wealth. In addition, we present an online investment algorithm, namely the mixture successive constant rebalanced portfolio with side-information, that relies on the mixture best constant rebalanced portfolio and side-information. The proposed online investment algorithm assumes the existence of an oracle, which, by exploiting the available side-information, is capable of predicting, with different levels of accuracy (Han and Kamber 2001), the state of the stock market for the next trading period. The empirical performance of the online investment algorithm is investigated using a set of numerical experiments concerning four major stock market data sets, namely the Dow Jones Industrial Average, the Standard and Poor’s 500, the Toronto Stock Exchange (Borodin et al. 2000) and the New York Stock Exchange (Cover 1991b; Helmbold et al. 1996). The results obtained emphasize the relevance of the proposed sequential investment strategy and underline the central role of the quality of the side-information in outperforming the best constant rebalanced portfolio. The remainder of the chapter is organized as follows. In Section 5.2 we introduce the notation and main definitions concerning the stock market, the price relative, the constant rebalanced portfolio and the successive constant rebalanced portfolio (Gaivoronski and Stella 2000). Side-information, the mixture best constant rebalanced portfolio and the mixture successive constant rebalanced portfolio are introduced and analysed in Section 5.3. Section 5.3 is concerned with the theoretical framework for online investment in the case where side-information is available and provides the theoretical analysis and comparison between the best constant rebalanced portfolio, the mixture best constant rebalanced portfolio and the mixture successive constant rebalanced portfolio. Finally, Section 5.4 presents and comments on the results of a set of numerical experiments concerning some of the main financial market data sets described in the specialized literature (Cover 1991b; Helmbold et al. 1996; Borodin et al. 2000).

5.2

CONSTANT REBALANCED PORTFOLIOS

Following Cover (1991b), a stock market vector is represented as a vector z ¼ ðz1 ; . . . ; zm Þ; such that zi ]0,
i 1, . . ., m, where m is the number of stocks and zi is the price relative, i.e. it represents the ratio of the price at the end of the trading period to the price at the beginning of the trading period.

88 j CHAPTER 5

A portfolio is described by the vector x ¼ ðx1 ; . . . ; xm Þ; such that ( x2X¼

x j xi  0; 8i ¼ 1; . . . ; m;

m X

) xi ¼ 1 :

i¼1

The portfolio x is an allocation of the current wealth across the stocks in the sense that xi represents the fraction of wealth invested in the ith stock. By assuming that x and z represent, respectively, the portfolio and the stock market vector for one investment period, the wealth relative (i.e. the ratio of the wealth at the end of the trading period to the wealth at the beginning of the trading period), given by S ¼ xT z; represents the factor by which the wealth increases/decreases in one investment period using portfolio x. The problem of portfolio selection consists of selecting a portfolio x that would maximize S in some sense. Financial theory has developed various notions of optimality for the portfolio selection problem. One possibility is to maximize the expected value of S subject to a constraint on the variance as proposed by the Sharpe
Markowitz theory of investment (Markowitz 1952), which deals with the long-term behaviour of fixed portfolios. However, the mean-variance investment framework does not take into proper account the possibilities of frequent portfolio rebalances, which are one of the most important features characterizing a stock market. To overcome this limitation, another possibility for the portfolio selection problem was proposed and described by Cover (1991a, b) that exploits the concept of the constant rebalanced portfolio (CRP), i.e. a portfolio such that, after each trading period, it is arranged in order to keep constant the fraction of wealth invested in every stock. By considering an arbitrary non-random sequence of n stock market vectors zð1Þ ; . . . ; zðnÞ , a CRP x achieves wealth /

Sðx; nÞ ¼

n Y

xT zðtÞ ;

t¼1

where we assume that the initial wealth (t 0) is normalized to 1 (S(x,0) 1). Within the class of constant rebalanced portfolios (CRPs), the best of such portfolios determined in hindsight, namely the best constant rebalanced portfolio (BCRP), i.e. the CRP computed by assuming perfect knowledge of future stock prices, possesses interesting properties. Indeed, Cover (1991a, b) showed that the wealth achieved by means of the

CONSTANT REBALANCED PORTFOLIOS AND SIDE-INFORMATION

j

89

BCRP is not inferior to that achieved by the best stock, to that associated with the value line and to that associated with the arithmetic mean. These properties have motivated increasing interest in the study and analysis of the main features of this investment strategy and the use of this portfolio as the reference benchmark to evaluate and compare sequential investment strategies.  Let us now formally introduce the BCRP xðnÞ that solves the following optimization problem: ð5:1Þ

max Sðx; nÞ: x2X

 ðnÞ

The vector x maximizes the wealth S(x,n) across the stock market vector sequence z(1), . . ., z(n) and therefore it is defined as the BCRP for the stock market vector sequence z(1), . . ., z(n).  ðnÞ The portfolio x cannot be used, however, for actual stock selection, at trading period n, because it explicitly depends on the sequence z(1), . . ., z(n) which becomes known only after the expiration of this time interval. Therefore, a reasonable objective might be to construct a sequence of portfolios fxðtÞ g, i.e. a sequential investment strategy such that, at trading periods t 2, . . ., n, the portfolio x(t), used for stock selection, depends on the sequence z(1), . . ., z(t1). Let us denote by S ({x(t)}, n) the wealth generated after n trading periods by successive application of the sequence of portfolios {x(t)}, then S



n   Y T xðtÞ ; n ¼ xðtÞ zðtÞ : t¼1

It would be desirable if such a sequential investment strategy {x(t)} would yield wealth in  ðnÞ some sense close to the wealth obtained by means of the BCRP x . One such strategy was proposed by Cover (1991b) under the name of the universal portfolio (UP) and consists of selecting the investment portfolio as follows: ð1Þ

b x

¼

 1 1 ;...; ; m m

R ðtþ1Þ

b x

¼ RX

xSðx; tÞ dx

Sðx; tÞ dx X

:

ð5:2Þ

The UP (5.2) has been shown by Cover (1991a) to possess a very interesting property: it has the same exponent, to first order, as the BCRP. Formally, by letting n  ðtÞ   Y T b x ðtÞ zðtÞ S b x ;n ¼ t¼1

be the wealth achieved by means of the UP, then it has been shown that 1 n

log S

 ðtÞ   1   ðnÞ  b x ; n log S x ; n ! 0; n

90 j CHAPTER 5

with the following inequality holding: S

  ðnÞ   ðtÞ   b x ; n  S x ; n Cn n ðm 1Þ=2 ;

where Cn tends to some limit along subsequences for which   ðnÞ  1   ðnÞ    ðnÞ  W x ; n ¼ log S x ; n ! W x n for some strictly concave function W(x) (see theorem 6.1 of Cover 1991a). Another example of an investment strategy that exploits the definition of BCRP has been proposed by Gaivoronski and Stella (2000). This strategy, called the successive constant rebalanced portfolio (SCRP), selects the investment portfolio as follows: ð1Þ

e x

¼

 1 1 ;...; ; m m

e x ðtþ1Þ ¼ arg max Sðx; tÞ; x2X

Pm where X ¼ fx : xi  0; 8i ¼ 1; . . . ; m; SCRP fe x ðtÞ g possesses interi¼1 xi ¼ 1g.    The esting properties. Indeed, its asymptotic wealth S e x ðtÞ ; n coincides with the wealth obtained by means of the BCRP to first order in the exponent, i.e.  ðtÞ   1   ðnÞ  1 log S e x ; n log S x ; n ! 0; n n with the following inequality holding: S

 ðtÞ     ðnÞ  2 e x ; n  S x ; n Cðn 1Þ 2K =d ;

ð5:4Þ

where K ¼ supt;x2X kHx ½lnðxT zðtÞ Þk, while C and d are constants.

5.3

ONLINE INVESTMENT WITH SIDE-INFORMATION

BCRP, UP and SCRP are investment strategies designed to deal with the portfolio selection problem in the case where no additional information is available concerning the stock market. However, it is common practice that investors, fund managers and private investors adjust their portfolios, i.e. rebalance, using various sources of information concerning the stock market, which can be conveniently summarized by the concept of side-information. A typical example of side-information originates from sophisticated trading strategies that often develop signalling algorithms that individuate the nature of the investment opportunity about to be faced. In this context, side-information is usually considered to be a causal function of past stock market performance. Therefore, the availability of sideinformation concerning the stock market calls for the definition of a new investment benchmark, other than the BCRP, capable of appropriately exploiting side-information about the stock market. In this direction, Cover and Ordentlich (1996) proposed the state

CONSTANT REBALANCED PORTFOLIOS AND SIDE-INFORMATION

j

91

constant rebalanced portfolio, which is capable of appropriately exploiting the available sideinformation concerning the stock market. The state constant rebalanced portfolio achieves, to first order in the exponent, the same wealth as the best side-information-dependent investment strategy determined in hindsight from the observed stock market vector and side-information outcomes. Following the main ideas of Cover and Ordentlich (1996) we propose and describe a theoretical framework for online investment in the case where side-information is available concerning the stock market. We propose a mathematical model for the stock market in the presence of side-information and define a new investment benchmark that appropriately exploits the available side-information. The proposed mathematical model assumes that, at any trading period t, the stock market can be in one of H possible states belonging to H  { 1, . . ., H }. The stock market state at trading period t influences the stock market vector z(t). Formally, the model assumes that the infinite sequence of stock market vectors zð1Þ ; . . . ; zð1Þ is a realization from a mixture consisting of H components. Each mixture’s component is associated with a given stock market state. Therefore, any infinite sequence of stock market vectors zð1Þ ; . . . ; zð1Þ partitions into H mutually exclusive subsets Z1 ; . . . ; ZH , each subset Zh containing those stock market vectors associated with the corresponding stock market state h. It should be emphasized that the partitioning Z1 ; . . . ; ZH is assumed to be data independent, i.e. it does not depend on the given infinite sequence of stock market vectors zð1Þ ; . . . ; zð1Þ , but it only depends on the mixture’s components. The same model applies to finite stock market vector sequences zð1Þ ; . . . ; zðnÞ . However, in order to avoid confusion, for any given finite stock market vector sequence zð1Þ ; . . . ; zðnÞ , we will let ðnÞ Zh be the subset containing only those stock market vectors z(t), t 1, . . ., n, associated ðnÞ ðnÞ with the stock market state h. Furthermore, we let nh be the cardinality of Zh , PH ðnÞ ðnÞ h¼1 nh ¼ n and assume that nh ! 1 as n 0 ,
h  H. This stock market model requires the definition of a new investment benchmark, the mixture best constant rebalanced portfolio (MBCRP), in the case where side-information is available, and can be used to make an inference about the current stock market state. To introduce the mathematical framework for dealing with the proposed stock market model and side-information, let us take into account the generic stock market state h and  ðnÞ let xh solve the optimization problem (5.1) for those stock market vectors zð1Þ ; . . . ; zðnÞ  ðnÞ ðnÞ belonging to subset Zh , i.e. xh maximizes the logarithmic wealth relative associated with stock market state h

log Sh ðx; nÞ ¼

n X t¼1

  IzðtÞ 2ZðnÞ log xT zðtÞ :

ð5:5Þ

h

 ðnÞ

Let us now expand (5.5) in a Taylor series, up to second order, centred at xh to obtain

92 j CHAPTER 5

"

  ðnÞ T ðtÞ   ðnÞ T ðtÞ  x xh z log Sh ðx; nÞ ¼ IzðtÞ 2ZðnÞ log xh z þ  ðnÞ T ðtÞ h xh z t¼1      ðnÞ T  ðnÞ  x xh zðtÞ ðzðtÞ ÞT x xh  ðnÞ 2 2 x h T zðtÞ #   ðnÞ T ðtÞ 3 x xh z þ 3  3 ET zðtÞ n X

 ðnÞ

for a given vector  between x and xh . Equation (5.6), by exploiting (5.5) and by isolating the contribution for each of its four terms, can be rewritten as  ðnÞ

log Sh ðx; nÞ log Sh ðxh ; nÞ "  ðnÞ T ðtÞ n X x xh z IzðtÞ 2ZðnÞ ¼ ðnÞ h x h T zðtÞ t¼1 #    ðnÞ T  ðtÞ ðtÞ T   ðnÞ   ðnÞ T ðtÞ 3 x xh z ðz Þ x xh x xh z : þ 3   ðnÞ 2 3 ET zðtÞ 2 x h T zðtÞ

ð5:7Þ

Now, according to Gaivoronski and Stella (2000), the following conditions hold: 

Asymptotic independence

lim inf gmin n

n X

IzðtÞ 2ZðnÞ h

t¼1

zðtÞ ðzðtÞ ÞT kzðtÞ k

!

2

 dh > 0;

where by gmin(A) we denote the smallest eigenvalue of matrix A;  Uniform boundedness ðtÞ

0 < z  zi  z þ ;

8t; i:

Theorem 5.2 in Gaivoronski and Stella (2000) ensures that (5.5) is strictly concave on X uniformly over n and therefore the following inequality holds:   ðnÞ T ðtÞ n    ðnÞ  X x xh z dh   ðnÞ 2  x x IzðtÞ 2ZðnÞ log Sh ðx; nÞ log Sh xh ; n   h  : ðnÞ h 2 x h T zðtÞ t¼1  ðnÞ

ð5:8Þ

Now, from optimality conditions, applied to the BCRP xh , we know that, for each stock market state h and for any CRP x, the following condition holds:

CONSTANT REBALANCED PORTFOLIOS AND SIDE-INFORMATION

n X

 IzðtÞ 2ZðnÞ h

t¼1

j

93

 ðnÞ T ðtÞ

x xh

ðnÞ x h T zðtÞ

z

 0;

ð5:9Þ

and, therefore, by combining (5.8) with (5.9) we can write    ðnÞ  d   ðnÞ 2  log Sh xh ; n log Sh ðx; nÞ  h x xh  : 2

ð5:10Þ ðnÞ

Finally, by recalling (5.8) and using the relationship between dh and nh , we obtain the inequality    ðnÞ  1 ðnÞ   ðnÞ 2  log Sh xh ; n log Sh ðx; nÞ  nh gh x xh  ; 2

ð5:11Þ

where gh is the lim inf of the minimum eigenvalue of the matrix n 1 X ðnÞ

nh

t¼1

zðtÞ zðtÞT IzðtÞ 2ZðnÞ  2 : h zðtÞ 

Let us now return to the proposed mathematical framework for dealing with stock markets and side-information, i.e. let us take into account the case where the stock market can be in one of H possible states belonging to H  {1, . . ., H}. In such a framework, given a sequence of stock market vectors zð1Þ ; . . . ; zðnÞ , where each stock market vector is associated with a given stock market state, the achieved logarithmic wealth, by means of  ðnÞ the investment strategy fxH g that, at each trading period t, exploits knowledge of the  ðnÞ current stock market state h and applies the corresponding BCRP xh according to (5.5), is given by

log S

H   ðnÞ    ðnÞ   X log Sh xh ; n ; xH ; n ¼

ð5:12Þ

h¼1

whereas the logarithmic wealth achieved by means of any CRP x is given by P log Sðx; nÞ ¼ nt¼1 logðxT zðtÞ Þ. The difference between the logarithmic wealth achieved by means of the investment  ðnÞ strategy fxH g and the logarithmic wealth achieved by means of any other CRP x is given by H n X   ðnÞ  X     ðnÞ   log Sh xh ; n log xT zðtÞ ; log S xH ; n log Sðx; nÞ ¼ h¼1

and from (5.11) we can write

t¼1

94 j CHAPTER 5

log S

H   ðnÞ   1 X ðnÞ   ðnÞ 2 xH ; n log Sðx; nÞ  nh gh x xh  : 2 h¼1  ðnÞ

 ðnÞ

Therefore, given the investment strategy fxH g and the BCRP x , the following inequality holds: log S

H   ðnÞ  ðnÞ    1X   ðnÞ   ðnÞ nh gh x xh 2 : xH ; n log S xðnÞ ; n  2 h¼1

ð5:13Þ

The bound (5.13) strongly motivates interest in investment strategies depending on knowledge of the current stock market state. Indeed, coming back to wealth, the following condition holds: ! H    ðnÞ   1X  ðnÞ  ðnÞ 2 ðnÞ   ðnÞ nh gh x xh  ; S xH ; n  Sðx ; nÞ exp 2 h¼1

ð5:14Þ

 ðnÞ

and, therefore, the investment strategy fxH g achieves a wealth that is exponential with  ðnÞ respect to the wealth achieved by means of the BCRP x .  ðnÞ Let us now give the formal definition of the MBCRP investment strategy fxH g, that, at each trading period t, exploits knowledge of the current stock market state h to select and  ðnÞ to apply the corresponding BCRP xh . Definition 5.3.1 (Mixture best constant rebalanced portfolio): Given a stock market characterized by the states in H  {1, . . ., H}, the MBCRP is defined as the following investment strategy, i.e. the set of BCRPs:   ðnÞ    ðnÞ  ðnÞ  xH ¼ x1 ; . . . ; xH ;

ð5:15Þ

 ðnÞ

where each portfolio xh is the BCRP, according to (5.5), associated with stock market state h. The MBCRP is the investment strategy that, at each trading period t, knows the stock  ðnÞ market state h and therefore invests using the corresponding BCRP xh . The MBCRP possesses an interesting property; it outperforms by an exponential factor the corresponding BCRP in terms of the wealth achieved as stated by the following theorem.  ðnÞ

Theorem 5.3.2: The wealth accumulated after n trading periods from the MBCRP fxH g  ðnÞ outperforms the wealth of the corresponding BCRP x by an exponential factor. Formally, ! H   ðnÞ    ðnÞ   1 X ðnÞ   ðnÞ  ðnÞ 2 n h g h x x h  : S xH ; n  S x ; n exp 2 h¼1

Proof:

The proof follows directly from Equation (5.14).

ð5:16Þ

I

CONSTANT REBALANCED PORTFOLIOS AND SIDE-INFORMATION

j

95

Let us now introduce an online investment strategy, namely the mixture successive constant rebalanced portfolio (MSCRP), which approximates the MBCRP and relies on the successive constant rebalanced portfolio (SCRP) online investment strategy introduced and analysed by Gaivoronski and Stella (2000). Definition 5.3.3 (Mixture successive constant rebalanced portfolio): Given a stock market characterized by the states in H  {1, . . ., H}, the MSCRP is defined as the following investment strategy, i.e. the set of SCRPs:  ðtÞ   ðnÞ   ðtÞ  e x1 ; . . . ; e xH ; xH ¼ e

ð5:17Þ

ðtÞ

where each portfolio fe xh g is the SCRP associated with the stock market state h at trading period t. The MSCRP is the investment strategy that, at each trading period t, knows the ðtÞ stock market state h and therefore invests using the corresponding SCRP e xh . The relationship between the accumulated wealth, after n trading periods, by means of the MBCRP benchmark and by means of the MSCRP online investment strategy, is clarified by the following theorem. Theorem 5.3.4: For each stock market characterized by states in H  {1, . . ., H}, the MSCRP (5.17) is universal with respect to the MBCRP (5.15). Formally,  ðnÞ   1   ðnÞ   1 log S e xH ; n log S xH ; n ! 0: n n Proof: From property (5.4) of the SCRP (5.3) online investment strategy, for any given stock market state h, the following inequality holds:  ðtÞ   ðhÞ  ðnÞ Q ðhÞ   ðnÞ  nh 1 xh ; n C ; S xh ; n  S e

ð5:18Þ

where C(h) and Q(h) are constants depending on the stock market state h. Then, by combining Equation (5.12) with Equation (5.18), it is possible to write the inequality

S

H H   ðnÞ  Y  ðtÞ   ðhÞ  ðnÞ Q ðhÞ   ðnÞ   Y nh 1 S xh ; n  S e xh ; n C ; xH ; n ¼ h¼1

h¼1

and, thus, by taking the logarithm of both sides we obtain

log S

H   ðnÞ   X   ðtÞ     ðnÞ  xH ; n  log S e xh ; n þ Q ðhÞ log nh 1 þ log C ðhÞ ; h¼1

ð5:19Þ

96 j CHAPTER 5

which can be rewritten as

log S

H  ðnÞ   X  ðhÞ    ðnÞ    ðnÞ  xH ; n  Q log nh 1 þ log C ðhÞ : xH ; n log S e h¼1

Therefore, dividing both sides by n, it is possible to conclude that  ðnÞ   1   ðnÞ   1 log S e xH ; n log S xH ; n ! 0; n n which completes the proof. I Let us now compare the MSCRP with the corresponding BCRP in terms of the accumulated wealth after n trading periods. Theorem 5.3.5: For each stock market characterized by the states in H{1, . . ., H}, the ðnÞ accumulated wealth after n trading periods by means of the MSCRP fe xH g outperforms the  accumulated wealth by means of the corresponding BCRP xðnÞ by an exponential factor. Formally, we have that, for n 0, the following inequality holds:

S

Proof:

H   ðnÞ   ðnÞ   ðnÞ  ðnÞ 2  ðnÞ   1X e nh gh x xh  xH ; n  S x ; n exp 2 h¼1 !    ðnÞ Q ðhÞ log nh 1 log C ðhÞ :

Inequality (5.16) from theorem 5.3.2 states that ! H    ðnÞ     ðnÞ  1X  ðnÞ 2 ðnÞ   ðnÞ S xH ; n  S x ; n exp nh gh x xh  ; 2 h¼1

therefore using (5.19) it is possible to write H Y  ðtÞ   ðhÞ  ðnÞ Q ðhÞ nh 1 S e xh ; n C h¼1

! H    ðnÞ    ðnÞ   1X  ðnÞ 2 ðnÞ   ðnÞ  ;  S xH ; n  S x ; n exp nh gh x xh 2 h¼1 and, by taking the logarithm and reordering, we obtain the inequality

CONSTANT REBALANCED PORTFOLIOS AND SIDE-INFORMATION

log S 

j

97

  ðnÞ   ðnÞ   e xH ; n log S x ; n H   ðnÞ  ðnÞ 2 1X ðnÞ nh gh x xh  2 h¼1



H X

 ðnÞ  Q ðhÞ log nh 1 log C ðhÞ :

h¼1

Now, by taking the limit n 0 , where for each stock market state h we assume that ðnÞ nh ! 1, we have   ðnÞ   ðnÞ   S e xH ; n  S x ; n exp 

ðnÞ

H    ðnÞ  ðnÞ 2 1X ðnÞ nh gh x xh  2 h¼1



Q ðhÞ log nh 1 log C ðhÞ



! :

Furthermore, we can write   ðnÞ  S x ; n exp

H    ðnÞ  ðnÞ 2 1X ðnÞ nh gh x xh  2 h¼1

!   ðnÞ Q ðhÞ log nh 1 log C ðhÞ 

! H    ðnÞ  1X  ðnÞ 2 ðnÞ   ðnÞ nh gh x xh  ;  S x ; n exp 2 h¼1 and therefore it is possible to write the following inequality: ! H   ðnÞ     ðnÞ  1X  ðnÞ 2 ðnÞ   ðnÞ S e xH ; n  S x ; n exp nh gh x xh  ; 2 h¼1 which completes the proof.

I

According to theorems 5.3.2, 5.3.4 and 5.3.5 the MSCRP investment strategy (definition 5.3.3) possesses interesting theoretical properties in the case where side-information is available concerning the stock market. However, the MSCRP cannot be used directly to invest in the stock market because, for each trading period t, it assumes perfect knowledge about the current stock market state h, which is indeed known just after the current trading period t has expired. Therefore, it would be desirable to develop an algorithm that, by using past information about the stock market, makes predictions about the stock

98 j CHAPTER 5

market state h for the next trading period t. More precisely, the algorithm to be developed should exploit the available side-information to provide predictions about the stock market state associated with the next trading period t. In order to clarify how the prediction task concerning the stock market state h for the next trading period t can be formulated and therefore to complete the proposed mathematical framework for online investment with side-information, we adopt the Bayesian paradigm. In particular, we let m 2 R be the side-information vector and p(Y|h) the state conditional probability density function for Y, i.e. the probability density function for the side-information vector Y conditioned on h being the stock market state. Furthermore, we let P(h) be the a priori probability that the stock market is in state h. Then, the posterior probability P(h|Y) that the stock market is in state h given the side-information vector Y can be computed using the celebrated Bayes’ rule (Duda and Hart 1973) as follows: PðhjY Þ ¼

pðY jhÞPðhÞ : pðY Þ

ð5:20Þ

The Bayes’ rule (5.20) offers a precious theoretical model for exploiting the available side-information and therefore to make an inference concerning the stock market state for the next trading period. However, the quantities on the right-hand side of (5.20), i.e. the state conditional probability density p(Y|h) as well as the a priori probability P(h), are unknown and must be estimated by combining a priori knowledge with the available data. Several computational approaches and algorithms have been proposed and developed in recent years and published in the specialized literature (Duda and Hart 1973; Bernardo and Smith 2000; Duda et al. 2001; Zaffalon 2002; Congdon 2003; Zaffalon and Fagiuoli 2003) to deal with the problem of the unknown likelihood p(Y |h) and prior P(h), and treatment of such problem is out of the scope of this chapter. In this chapter we assume the existence of an oracle that, at each trading period t and using the available side-information vector Y, is capable of making an inference concerning the true stock market state for trading period t  1 with different accuracy levels q (Han and Kamber 2001). In each trading period t the oracle first accesses the available sideinformation vector Y and then makes inference hb concerning the stock market state h for the next trading period t  1. The oracle is assumed to provide predictions with different levels of accuracy, i.e. it is assumed that the following condition holds: Pðhb ¼ hÞ ¼ q;

8h 2 H:

ð5:21Þ

It is worth noting that condition (5.21) states that the oracle is capable of providing predictions that, in the long run, are associated with a misclassification error equal to 1q with q defined as the oracle’s accuracy level. Let us now describe an online investment algorithm that relies on the MSCRP investment strategy and which, to make an inference concerning the stock market state for each trading period, uses the predictions provided by means of the considered oracle.

CONSTANT REBALANCED PORTFOLIOS AND SIDE-INFORMATION

j

99

Algorithm 5.1 MSCRP with side-information 1. At the beginning of the first trading period take ð1Þ

e x

¼

 1 1 ;...; : m m

2. At the end of trading period t 1, . . . use side-information Y to make an inference concerning the stock market state at trading period t  1. Assume the oracle makes an inference in favour of stock market state h. 3. Let zð1Þ ; . . . ; zðtÞ be the stock market vectors available up to trading period t. Compute e x ðtþ1Þ as the solution of the following optimization problem: max log Sh ðx; tÞ; x2X

Pt where log Sh ðx; tÞ ¼ s¼1 IzðsÞ 2ZðtÞ logðxT zðsÞ Þ: h The MSCRP with side-information algorithm exploits the side-information Y available at trading period t and makes inference hb ¼ h for the next stock market state, i.e. at trading period t  1. Then, the MSCRP with side-information algorithm exploits the ðtÞ forecast hb ¼ h and invests in the stock market by means of the corresponding SCRP e xh . Therefore, the MSCRP with side-information algorithm approximates the MSCRP sequential investment strategy. The theoretical properties of the MSCRP with sideinformation algorithm clearly depend on the capability of correctly assessing the stock market state for the next trading period and therefore depend on the oracle’s accuracy level q (5.21).

5.4

NUMERICAL EXPERIMENTS

This section illustrates and comments on the results obtained from a set of numerical experiments concerning the MSCRP with side-information algorithm in the case where four major stock market data sets described in the specialized literature are considered. The four stock market data sets are the Dow Jones Industrial Average, the Standard and Poor’s 500, the Toronto Stock Exchange (Borodin et al. 2000) and the New York Stock Exchange (Cover 1991b; Helmbold et al. 1996). The first data set consists of the 30 stocks belonging to the Dow Jones Industrial Average (DJIA) for the 2-year period (507 trading periods, days) starting from Jan 14, 2001 to Jan 14, 2003. The second data set consists of the 25 stocks from the Standard and Poors 500 (S&P500) having the largest market capitalization during the period starting from Jan 2, 1998 to Jan 31, 2003 (1276 trading periods, days). The Toronto Stock Exchange (TSE) data set consists of 88 stocks for the 5-year period (1259 trading periods, days) starting from Jan 4, 1994 to Dec 31, 1998. Finally, the New York Stock Exchange (NYSE) data set consists of 5651 daily prices (trading periods) for 35 stocks for the 22year period starting from July 3, 1962 to Dec 31, 1984.

100 j CHAPTER 5

We assume that each stock market is characterized by two states (H  {1, 2}) with state h  1 associated with trading periods for which the average value of price relatives over all stocks is greater than one, whereas state h  2 is associated with the remaining trading periods. This partitioning can be conveniently interpreted as generating bull (h  1) and bear (h  2) trading periods. It represents a first proposal and much effort will be devoted to finding alternative partitioning criteria to improve the effectiveness of the MSCRP with side-information online investment algorithm. Numerical experiments were performed without the help of any inference device in the sense that no computational devices were used to predict the stock market state h for the next trading period. The inference concerning the stock market state h for the next trading period is obtained by exploiting predictions provided by means of an oracle associated with different accuracy levels q (5.21). It should be noted that an accuracy level equal to one (q  1) corresponds to perfect knowledge of the stock market state h before the investment step takes place, thus leading ðtÞ to invest according to the corresponding optimal SCRP fe xh g. The results associated with the numerical experiments in the case where the oracle is assumed to have perfect knowledge of the stock market state (q  1) are reported in Table 5.1. The data reported in Table 5.1 demonstrate the effectiveness of the MSCRP with side-information: indeed, the achieved wealth significantly outperforms that achieved by means of the corresponding BCRP algorithm. However, it is unrealistic to assume perfect knowledge concerning the stock market state and a further investigation is required. It is of central relevance to analyse how the MSCRP with side-information behaves in the case where different values of the accuracy level q are considered. To this extent, a set of numerical experiments for the DJIA, S&P500, TSE and the NYSE stock market data sets was planned and performed by considering different values for the oracle’s accuracy level q. The numerical experiments were organized in such a way that, at each trading period, given the oracle’s accuracy level q, we extract a random number r from a uniform distribution in the interval [0,1]. If the extracted random number r is less than or equal to the selected oracle’s accuracy level q, i.e. if r 5 q, then the stock market state h for the next trading period is correctly predicted, otherwise the stock market state is wrongly assessed. The results are summarized for each stock market data set using the mean value of achieved wealth by means of the MSCRP with side-information. The mean value of the wealth is computed using the described random sampling procedure in the case where 1000 samples are extracted for each trading period and different values of the oracle’s accuracy

TABLE 5.1 Wealth for the MSCRP with Side-Information Assuming Perfect Knowledge (q  1)

DJIA S&P500 TSE NYSE

BCRP

MBCRP

MSCRP

1.24 4.07 6.78 250.59

54.70 10 527.00 591.61 3.62E10

21.27 3750.00 45.08 1.02E10

CONSTANT REBALANCED PORTFOLIOS AND SIDE-INFORMATION

j

101

level q belonging to the interval [0,1]. The stock market state frequency, i.e. the percentage of trading periods associated with each stock market state h, is reported in Table 5.2. The results of the numerical experiments are summarized using a graphical representation of the extra wealth obtained using the MSCRP with side-information over the corresponding BCRP, defined as ðnÞ

o¼

Sðfe xH g; nÞ ðnÞ

Sðx ; nÞ

ð5:22Þ

;

with respect to the oracle’s accuracy level q and the trading period t for each data set (Figures 5.1
5.4). Figures 5.1 to 5.4, where the minimum value of v (5.22) is zero, clearly show the exponential nature of the extra wealth obtained for the MSCRP with side-information over the corresponding BCRP. It is possible to observe the exponential nature of the extra wealth v for an increasing number t of trading periods and depending on the oracle’s accuracy level q. The oracle’s accuracy level required to achieve extra wealth v (5.22) equal to one, i.e. the value of q such that the MSCRP with side-information achieves the same wealth as the /

TABLE 5.2 Stock Market State Frequency

DJIA S&P500 TSE NYSE

h 1

h 2

0.48 0.52 0.58 0.53

0.52 0.48 0.42 0.47

DJIA

15

ω

10

5

500 400 300 200 t

100 0

FIGURE 5.1 MSCRP extra wealth for the DJIA data set.

0.2

0.4

0.6 q

0.8

1

102 j CHAPTER 5 S&P500

800

ω

600 400 200

1200 1000 800

600 400 t 200

0

0.2

0.4

0.6

0.8

1

q

FIGURE 5.2 MSCRP extra wealth for the S&P500 data set.

TSE

7 6

ω

5 4 3 2 1 1200 1

1000 800

0.8 600

t

0.6 0.4

400 200

0.2

q

0

FIGURE 5.3 MSCRP extra wealth for the TSE data set.

corresponding BCRP, is defined as the oracle’s parity accuracy level and is reported, for each stock market data set, in Table 5.3. The data reported in Table 5.3 allow us to assess the effectiveness of the MSCRP with side-information depending on the oracle accuracy level q required. It is interesting to investigate the relationship between the extra wealth v and the variables: number of

CONSTANT REBALANCED PORTFOLIOS AND SIDE-INFORMATION

j

103

NYSE

×107 5

ω

4 3 2 1

5000 4000 3000 t

2000 1000

0.2

0

0.6 q

0.4

1

0.8

FIGURE 5.4 MSCRP extra wealth for the NYSE data set. TABLE 5.3 Oracle’s Parity Accuracy Levels q DJIA S&P500 TSE NYSE

0.60 0.59 0.77 0.55

TABLE 5.4 Parameter Estimates for Model (5.23) and R2 Values

DJIA S&P500 TSE NYSE

a0

a1

a2

a3

R2

0.4272 1.0097 0.3058 3.8365

0.0086 0.0084 0.0056 0.0041

0.9871 1.8281 0.8444 7.7298

0.0148 0.0144 0.0075 0.0073

0.9791 0.9850 0.9814 0.9880

N 51 128 127 570

207 876 159 751

trading periods t and oracle accuracy level q. To this end the linear regression model for the logarithm of the extra wealth v with four d.o.f., logðoÞ ¼ a0 þ a1 t þ a2 q þ a3 tq;

ð5:23Þ

was fitted using a different number of data points N for each stock market data set. Parameter estimates for model (5.23) together with the corresponding R2 values are reported in Table 5.4. Finally, Figures 5.5
5.8 plot the estimated value, using regression model (5.23), of the logarithm of the extra wealth log(v) with respect to trading period t and oracle accuracy level q. /

104 j CHAPTER 5 DJIA

2

Log (ω)

1 0 –1 –2 –3 500 400 300 t 200

100 0

0.2

0.4

0.6 q

0.8

1

FIGURE 5.5 Estimated logarithm of the extra wealth for the DJIA data set.

S&P500

Log (ω)

5

0

–5

1200 1000

800 t

600

400

200

0

0.2

0.4

0.6 q

0.8

1

FIGURE 5.6 Estimated logarithm of the extra wealth for the S&P500 data set.

5.5

CONCLUSIONS AND FURTHER RESEARCH DIRECTIONS

This chapter deals with online portfolio selection in the case where side-information is available concerning the stock market. Its theoretical achievements strongly support the further study and investigation of the class of MBCRP investment strategies. The experimental evidence that the MSCRP with side-information investment algorithm outperforms, in terms of the achieved wealth, the corresponding BCRP investment algorithm by an exponential factor is a basic result of the present work. However, this empirical achievement strongly depends on the oracle’s accuracy level. Therefore, the

CONSTANT REBALANCED PORTFOLIOS AND SIDE-INFORMATION

j

105

TSE

Log (ω)

0 –2 –4 –6 1200 1000

800 t

600

400

200

0

0.2

0.4

0.6 q

0.8

1

FIGURE 5.7 Estimated logarithm of the extra wealth for the TSE data set.

NYSE

20

Log (ω)

10 0 –10 –20 5000 4000 3000 t 2000

1000 0

0.2

0.4

0.6 q

0.8

1

FIGURE 5.8 Estimated logarithm of the extra wealth for the NYSE data set.

problem is shifted to the study and development of efficient and reliable computational devices for predicting the stock market state, at each trading period, by exploiting the available side-information. A second issue of interest, which probably has significant interplay with the development of reliable prediction models, is the choice of the stock market states. These motivations orient the next step of this research to the study and analysis of stock market state partitioning criteria as well as to the study and development of efficient models for the prediction of the stock market state by exploiting the available side-information.

106 j CHAPTER 5

ACKNOWLEDGEMENTS The authors are grateful to the anonymous referees whose insightful comments enabled us to make significant improvements to the chapter. Special thanks go to Professor Georg Pflug for his comments and suggestions concerning the mathematical notation, which allowed us to significantly improve the clarity of the chapter.

REFERENCES Algoet, P.H. and Cover, T.M., Asymptotic optimality and asymptotic equipartition properties of log-optimum investment. Ann. Probabil., 1988, 2, 876
/898. Auer, P. and Warmuth, M.K., Tracking the best disjunction, in 36th Annual Symposium on Foundations of Computer Science., 1995, pp. 312
/321. Bell, R. and Cover, T.M., Competitive optimality of logarithmic investment. Math. Oper. Res., 1980, 5, 161
/166. Bernardo, J.M. and Smith, A.F.M., Bayesian Theory, 2000 (Wiley: New York). Blum, A. and Kalai, A., Universal portfolios with and without transaction costs. Mach. Learn., 1998, 30, 23
/30. Borodin, A., El-Yaniv, R.. and Gogan, V., On the competitive theory and practice of portfolio selection, in Proceedings of the Latin American Theoretical Informatics (Latin), 2000. Browne, S., The return on investment from proportional portfolio strategies. Adv. Appl. Probabil., 1998, 30, 216
/238. Congdon, P., Applied Bayesian Modelling, 2003 (Wiley: New York). Cover, T.M., Elements of Information Theory, 1991a (Wiley: New York); Chapter 15. Cover, T.M., Universal portfolios. Math. Finan., 1991b, 1, 1
/29. Cover, T.M. and Ordentlich, E., Universal portfolios with side information. IEEE Trans. Inf. Theory, 1996, 42, 348
/363. Duda, R.O. and Hart, P.E., Pattern Classification and Scene Analysis, 1973 (Wiley: New York). Duda, R.O., Hart, P.E. and Stork, D.G., Pattern Classification, 2001 (Wiley-Interscience: New York). Evstigneev, I.V. and Schenk-Hoppe`, K.R., From rags to riches: On constant proportions investment strategies J. Theor. Appl. Finan., 2002, 5, 563
/573. Gaivoronski, A. and Stella, F., A stochastic nonstationary optimization for finding universal portfolios. Ann. Oper. Res., 2000, 100, 165
/188. Gaivoronski, A. and Stella, F., On-line portfolio selection using stochastic programming. J. Econ. Dynam. Contr., 2003, 27, 1013
/1014. Han, J. and Kamber, M.T., Data Mining: Concepts and Techniques, 2001 (Morgan-Kaufmann: San Francisco, CA). Helmbold, D.P., Schapire, R.E., Singer, Y. and Warmuth, M.K., On-line portfolio selection using multiplicative updates, in International Conference on Machine Learning., 1996, pp. 243
/251. Herbster, M. and Warmuth, M., Tracking the best expert. 12th International Conference on Machine Learning., 1995, pp. 286
/294. Markowitz, H., Portfolio selection. J. Finan., 1952, 7, 77
/91. Singer, Y., Switching portfolios. J. Neural Syst., 1997, 8, 445
/455. Vovk, V. and Watkins, C., Universal portfolio selection, in 11th Annual Conference on Computational Learning Theory, 1998, pp. 12
/23. Zaffalon, M., The naive credal classifier. J. Statist. Plann. Infer., 2002, 105, 5
/21. Zaffalon, M. and Fagiuoli, E., Tree-based credal networks for classification. Reliable Comput., 2003, 9, 487
/509.

CHAPTER

6

Improving Performance for Long-Term Investors: Wide Diversification, Leverage and Overlay Strategies JOHN M. MULVEY, CENK URAL and ZHUOJUAN ZHANG CONTENTS 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Fixed-Mix Portfolio Models and Rebalancing Gains . . . . . . . . . . . . . . 6.3 Empirical Results with Historical Data and Fixed-Mix Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4 A Stochastic Programming Planning Model . . . . . . . . . . . . . . . . . . . . 6.5 Summary and Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Appendix 6.A: Mathematical Model for the Multi-Stage Stochastic Program

6.1

. . 107 . . 110 . . . . .

. . . . .

113 118 121 124 125

INTRODUCTION

A

LTERNATIVE ASSETS, INCLUDING HEDGE FUNDS, private equity and venture capital, have gained in popularity due to the low return of equities since early 2000 and the commensurate search by institutional investors to improve performance in order to regain lost funding surpluses. In theory, alternative assets possess small dependencies with traditional assets such as stocks, bonds and the general level of economic activity—GDP, earnings and interest rates. However, a number of alternative asset categories have shown greater dependencies than originally perceived. We focus on a special class of dynamic (multi-stage) strategies for improving performance in the face of these issues. To set the stage for our analysis, we describe the historical returns and risks of two prototypical benchmarks for long-term investors in the United States: (1) 70) S&P 500

107

108 j CHAPTER 6

and 30) long government bonds (70/30); and (2) 60) S&P 500 and 40) government bonds (60/40). These two strategies have proven resilient over the last five decades with reasonable performance during most economic conditions. We conduct our historical evaluation over the recent 12-year period—1 January 1994 to 31 December 2005.1 This interval is divided into two distinct periods—the first six years (high equity  23.5) annual returns), and the last six years (low equity  1.1) annual returns). For the tests, we employ the so-called fixed-mix rule; the portfolio composition is rebalanced to the target mix (either 70/30 or 60/40) at the beginning of each time period (monthly). We will show the benefits of this rule in subsequent sections. Unfortunately, both the 70/30 and 60/40 strategies greatly under performed their longterm averages in the second six-year period 2000 2005, leading to a massive drop in the surpluses of pension plans and other institutional investors. To evaluate performance, we employ two standard measures of risks—volatility, and maximum drawdown. Other measures—value at risk, conditional/tail value at risk, and certainty equivalent returns—are closely related to these. Risk-adjusted returns are indicated by Sharpe and return to drawdown ratios. As a significant issue, the under performance of the fixed-mix 70/30 and 60/40 benchmarks as well as similar approaches over the past six years has caused severe difficulties for institutional investors, especially pension trusts. The large drawdown values—28.9) and 22.8), respectively, during 2000 2005 pinpoint the problems better than volatility—10.4) and 8.9), respectively, for the two benchmarks (Table 6.1). How can an investor improve upon these results? First, she might discover assets that provide higher returns than either the S&P 500 or long government bonds. Categories such as real estate investment trusts (REIT) have done just that over the past decade.2 Investors continue to search for high performing assets. Once a set of asset categories is chosen, there is a decision regarding the best asset allocation. Much has been written about financial optimization models. Rather than performing an optimization, we can search for novel diversifying assets. In this case, we might accept equal or lower expected returns in exchange for an improved risk profile. To this end, investors interested in achieving wide diversification might turn to assets such as foreign equity, emerging market equity and debt, and so on. Given wide diversification, we can apply leverage to achieve higher returns and lower (or equal) risks than the 70/30 or the 60/40 mix. Mulvey (2005) discusses increasing diversification and associated leverage for improving risk adjusted returns. Alternatively, a savvy investor might dynamically modify her asset mix as conditions warrant—moving into equity when certain indictors are met or reducing equity exposure when other conditions occur. Such an investor applies more complex decision rules than fixed-mix. Numerous fundamental and technical approaches are employed in this quest. In Section 6.4, we show that many dynamic strategies can be incorporated within the context of a multi-stage stochastic program. /

/

1 2

Performance results for several alternative asset categories are unavailable or suspect before 1994. As always, these assets are not guaranteed to provide superior results in the future.

IMPROVING PERFORMANCE FOR LONG-TERM INVESTORS

j

109

TABLE 6.1 Historical Returns for 70/30 and 60/40 Fixed-Mix Benchmarks and Variants (Large Drawdowns Are Common over the Period)

1994 /2005

1994 /1999

2000 /2005

S&P 500 LB Agg. S&P equal weighted index LB 20-year STRIPS index Geometric return Standard deviation Sharpe ratio Maximum drawdown Return/drawdown Geometric return Standard deviation Sharpe ratio Maximum drawdown Return/drawdown Geometric return Standard deviation Sharpe ratio Maximum drawdown Return/drawdown

100) 100)

10.5) 14.8) 0.45 44.7) 0.24 23.5) 13.6) 1.37 15.4) 1.53 1.1) 15.2) 0.25 44.7) 0.03

6.8) 4.5) 0.66 5.3) 1.29 5.9) 4.0) 0.24 5.2) 1.14 7.7) 5.1) 0.99 5.3) 1.47

70) 30)

9.7) 10.4) 0.56 28.9) 0.33 18.2) 10.0) 1.32 10.2) 1.77 1.8) 10.4) 0.09 28.9) 0.06

60) 40)

9.3) 9.1) 0.61 22.8) 0.41 16.4) 8.9) 1.30 8.5) 1.92 2.7) 8.9) 0.01 22.8) 0.12

70) 30) 12.5) 11.5) 0.76 15.5) 0.81 14.4) 11.2) 0.85 11.8) 1.22 10.6) 11.8) 0.67 15.5) 0.69

60) 40) 12.4) 10.8) 0.80 11.6) 1.07 13.5) 10.8) 0.79 11.6) 1.16 11.3) 10.8) 0.79 10.4) 1.09

The developed ‘overlay’ securities/strategies prove beneficial for both fixed-mix and dynamic mix investors. To define an overlay in a simplified setting, we start with a singleperiod static model. (A multi-stage version appears in the Appendix.) An un-levered, long-only portfolio model allocates the investor’s initial capital, C, to a set of assets {I} via decision variables xi ]0 so as to optimize the investor’s random wealth at the horizon: ½SP Subject to

Maximize U ðZ1 ; Z2 Þ; e and Z2  RiskðwÞ: e Z1  EðwÞ X

xi ¼ C;

ð6:1Þ

i2I

e¼ w

X i2I

rei xi :

ð6:2Þ

The generic utility function U ðÞ consists of two terms—expected return and a risk function. The latter encompasses most implemented approaches, including volatility, downside, value-at-risk, conditional-value-at-risk and expected von Neumann Morgenstern utility (Bell 1995). Random asset returns are identified as rei . It is a simple matter to address traditional leverage: we add a borrowing variable y ]0 and replace Equations (6.1) and (6.2) with X i2I

xi C þ y;

ð6:3Þ

110 j CHAPTER 6

e¼ w

X

rei xi  reb y;

ð6:4Þ

where the borrowing rate is reb , with perhaps an upper limit on the amount of borrowing y 5uy . To improve risk-adjusted performance, the borrowing rate needs to be low enough so the optimal dual variable on the y ]0 constraint equals zero (i.e. the constraint is non-binding). In contrast to traditional leverage, the overlays (called securities, positions, assets) do not require a capital outlay. For example, two creditworthy investors might establish a forward contract on currencies between themselves. Herein, the net returns—positive or negative—are simply added to the investor’s horizon wealth. Under selective conditions, futures markets approximate this possibly favourable environment. In the static model, we expand the decision variables to include the overlays, xj ]0 for j  J. The relevant constraints, replacing (6.3) and (6.4), are (6.1) and e¼ w

X i2I

rei xi þ

X

 1Þxj ; ðrj g

ð6:5Þ

j2J

Importantly, due to the nature of futures markets, the overlay variables can refer to a wide variety of underlying strategies—long-only, short-only, or long short.3 Thus, the overlay variables xj ]0 for j  J indicate the presence of a particular futures market contract (long, short or dynamic strategy), and its size. In this chapter, we evaluate the overlays within a classical trend-following rule (Mulvey et al. 2004); alternative rules are worthy of future tests. For risk management purposes, we limit at each time period the designated notional value of the overlays to a small P multiple—say 1 5m54—of investor’s capital: j2J xj m C. Since capital is not directly allocated for the overlays, the resulting portfolio problem falls into the domain of risk allocation/budgeting. The static portfolio model may be generalized in a manner to multi-stage planning models (Appendix). However, as we will see, some of the standard features of asset performance statistics must be re-evaluated in a multi-stage environment. /

6.2

FIXED-MIX PORTFOLIO MODELS AND REBALANCING GAINS

The next three sections take up multi-stage investment models via fixed-mix rules. First, we discuss general issues relating to the fixed-mix rule; then we measure the advantages of the overlays for improving performance within a fixed-mix context. To start, we describe the advantages of fixed-mix over a static, buy-and-hold approach. The topic of re-balancing gains (also called excess growth or volatility pumping) as derived from the fixed-mix decision rule is well understood from a theoretical perspective. The fundamental solutions were developed by Merton (1969) and Samuelson (1969) for long-term investors. Further work was done by Fernholz and Shay (1982) and Fernholz 3 An overlay asset must include a form of investment strategy, since the investment must be re-evaluated before or at the expiration date of the futures or forward contract.

IMPROVING PERFORMANCE FOR LONG-TERM INVESTORS

j

111

(2002). Luenberger (1997) presents a clear discussion. We illustrate how rebalancing the portfolio to a fixed-mix creates excess growth. Suppose that the stock price process Pt is lognormal so that it can be represented by the equation dPt ¼ aPt dt þ sPt dzt ;

ð6:6Þ

where a is the rate of return of Pt and s2 is its variance, zt is Brownian motion with mean 0 and variance t. The risk-free asset follows the same price process with rate of return equal to r and standard deviation equal to 0. If we represent the price process of risk-free asset by Bt , dBt ¼ rBt dt:

ð6:7Þ

When we integrate (6.6), the resulting stock price process is 2

Pt ¼ P0 eðas =2Þtþszt :

ð6:8Þ

Clearly, the growth rate g :¼ a  s2 =2 is the most relevant measure for long-run performance. For simplicity, we assume equality of growth rates across all assets. This assumption is not required for generating excess growth, but it makes the illustration easier to understand. Let us assume that the market consists of n stocks with stock price processes P1;t ; . . . ; Pn;t each following the lognormal price process. A fixed-mix portfolio has a wealth process Wt that can be represented by the equation dWt Wt

¼

Z1 dP1;t P1;t

þ

þ

Zn dPn;t Pn;t

;

ð6:9Þ

where Z1 ; . . . ; Zn are the fixed weights given to each stock (proportion of capital allocated to each stock) which sum up to one: n X

Zi ¼ 1:

ð6:10Þ

i¼1

The fixed-mix strategy in continuous time always applies the same weights to stocks over time. The instantaneous rate of return of the fixed-mix portfolio at anytime is the weighted average of the instantaneous rates of returns of the stocks in the portfolio. In contrast, a buy-and-hold portfolio is one where there is no rebalancing and therefore the number of shares for each stock does not change over time. This portfolio can be represented by the wealth process Wt following dWt ¼ m1 dP1;t þ þ mn dPn;t ; where m1 ; . . . ; mn depicts the number of shares for each stock.

ð6:11Þ

112 j CHAPTER 6

Again for simplicity, let us assume that there is one stock and a risk-free instrument in the market. This case is sufficient to demonstrate the concept of excess growth in a fixedmix portfolio as originally presented in Fernholz and Shay (1982). Assume that we invest h portion of our wealth in the stock and the rest (1h) in the risk-free asset. Then the wealth process Wt with these constant weights over time can be expressed as dWt Wt

¼

ZdPt Pt

þ

ð1  ZÞdBt Bt

ð6:12Þ

;

where Pt is the stock price process and Bt is the risk-free asset value. When we substitute the dynamic equations for Pt and Bt , we get dWt Wt

¼ ðr þ Zða  rÞÞdt þ Zsdzt :

ð6:13Þ

As before, we assume the growth rate of the stock and the risk-free asset are equal. Hence a  s2 =2 ¼ r:

ð6:14Þ

From Equation (6.13), we can see that the rate of return of the portfolio, av, is aw ¼ r þ Zða  rÞ:

ð6:15Þ

From (6.14) this rate of return is equal to aw ¼ r þ Zs2 =2:

ð6:16Þ

The variance of the resulting portfolio return is s2w ¼ Z2 s2 :

ð6:17Þ

Hence the growth rate of the fixed-mix portfolio becomes gw ¼

aw  s2w 2

¼rþ

ðZ  Z2 Þs2 2

:

ð6:18Þ

This quantity is greater than r for 0 B h B 1. As it is greater than r, which is the growth rate of individual assets, the portfolio growth rate has an excess component, which is ðZ  Z2 Þs2 =2. Excess growth is due to rebalancing the portfolio constantly to a fixed-mix. The strategy moves capital out of stock when it performs well and moves capital into stock when it performs poorly. By moving capital between the two assets in the portfolio, a higher growth rate than each individual asset is achievable. See Dempster et al. (2007) for a more general discussion of this phenomenon. The buy-and-hold investor with equal returning assets lacks the excess growth component. Therefore, buy-and-hold portfolios will under-perform fixed-mix portfolios

IMPROVING PERFORMANCE FOR LONG-TERM INVESTORS

j

113

in various cases. We can easily see from (6.16) that the excess growth component is larger when s takes a higher value. In this sense, the volatility of an asset is considered not as a risk but rather can be an opportunity to create excess growth in a re-balanced portfolio although of course from (6.17) portfolio return volatility also scales with s versus s2 for rate of return. Surprisingly perhaps, there is no need for mean reversion in the stock price processes. Higher performance is obtained through greater volatility in individual assets. Accordingly as we will see in the next section, Sharpe ratios may not provide adequate information about the marginal impact of including an asset category within a fixed-mix, re-balanced portfolio.

6.3

EMPIRICAL RESULTS WITH HISTORICAL DATA AND FIXED-MIX STRATEGIES

In this section, we describe the results of applying several fixed-mix decision rules to data over the 12-year historical period summarized in Table 6.1. The purpose of these empirical tests is to set benchmarks, to find suitable mixes of assets, and to illustrate the advantages of the overlay variables, as compared with solely traditional assets. Of course dynamic decision rules such as the multi-stage stochastic programs discussed in Section 6.5, may be implemented in practice. Here again, the overlays prove to be beneficial for improving risk adjusted returns. In the historical results the portfolio is re-balanced monthly via the fixed-mix rule. First, we show that re-balancing gains were readily attainable over the turbulent period 1994 to 2005 by deploying assets so as to attain wide diversification and leverage. As described above, we follow the fixed-mix strategy—re-balancing the portfolio at the beginning of each month. The strategies work best when the investor incurs small transaction costs such as for tax-exempt and tax-deferred accounts. Index funds and exchange traded funds present ideal securities since they are highly liquid and can be moved with minimal transaction costs. Table 6.2 depicts the returns and volatilities for a set of 12 representative asset categories—both traditional and alternative—over the designated 12-year period—1994 to 2005. Annual geometric returns and the two risk measure values are shown. We focus on general asset categories rather than sub-categories, such as, small/medium/large equities, in order to evaluate general benefits. Clearly, further diversification is possible via other investment categories. Over the period, annual returns range from low2.6) (for currencies) to high 13.1) (for real estate investment trusts, REITs). Many assets display disparate behaviour over the two six-year sub-periods: The Goldman Sachs commodity index (GSCI) and NAREIT had their worst showing during 1994 1999—the lowest returns and highest drawdown values, whereas EAFE and S&P 500 had the opposite results. As a general observation, investors should be ready to encounter sharp drops in individual asset categories. Drawdown for half of the categories lies in the range 26) to 48) (Table 6.2). Two of the highest historical Sharpe ratios occur in the hedge fund categories: (1) the CSFB hedge fund index (0.87); and (2) the Tremont long/short index (0.78). In both cases, returns are greater than the S&P 500 index with much lower volatility. This performance /

23.5) 13.6) 1.37 15.4) 1.53

1.1) 15.2) 0.25 44.7) 0.03

1994 /1999 Return Std Dev Sharpe R. M. drawdown Ret/drawdown

2000 /2005 Return Std Dev Sharpe R. M. drawdown Ret/drawdown

7.7) 5.1) 0.99 5.3) 1.47

5.9) 4.0) 0.24 5.2) 1.14

6.8) 4.5) 0.66 5.3) 1.29

LB Agg.

1.5) 15.1) 0.08 47.5) 0.03

12.3) 13.8) 0.54 15.0) 0.82

6.8) 14.5) 0.20 47.5) 0.14

EAFE

2.7) 0.5) 0.00 0.0) N/A

4.9) 0.2) 0.00 0.0) N/A

3.8) 0.5) 0.00 0.0) N/A

T-bills

20.0) 13.9) 1.24 15.3) 1.31

6.5) 12.0) 0.13 26.3) 0.25

13.1) 13.1) 0.71 26.3) 0.50

NAREIT

15.7) 22.5) 0.57 35.4) 0.44

4.7) 17.4) 0.01 48.3) 0.10

10.0) 20.1) 0.31 48.3) 0.21

GSCI

7.4) 5.1) 0.92 7.7) 0.96

14.1) 9.9) 0.93 13.8) 1.02

10.7) 7.9) 0.87 13.8) 0.77

Hedge Fund index

7.2) 12.7) 0.35 13.9) 0.52

5.5) 11.5) 0.05 17.7) 0.31

6.4) 12.1) 0.21 17.7) 0.36

Man. Fut. index

5.1) 7.0) 0.34 15.3) 0.33

0.1) 6.7) 0.73 20.4) 0.00

2.6) 6.8) 0.18 28.7) 0.09

Currency index

5.6) 8.6) 0.34 15.0) 0.37

18.5) 11.6) 1.18 11.4) 1.62

11.9) 10.3) 0.78 15.0) 0.79

Tremont L/S

8.1) 17.0) 0.31 30.3) 0.27

17.1) 13.7) 0.89 19.9) 0.86

12.5) 15.4) 0.56 30.3) 0.41

S&P EWI

14.0) 16.3) 0.69 19.1) 0.73

7.2) 14.7) 0.16 22.8) 0.32

10.6) 15.5) 0.43 22.8) 0.46

20-year Strips

S&P 500  Standard and Poor 500 index, LB Bond  Lehman long aggregate bond index, EAFE  Morgan Stanley equity index for Europe, Australia, and the Far East, NAREIT  National Association of Real Estate Investment Trusts, GSCI  Goldman Sachs commodity index, Hedge Fund Index  Tremont hedge fund aggregate index, Man. Fut. Index  Tremont managed futures index, Tremont L/S  Tremont long/short equity index, S&P 500 equal weighted  Rydex S&P 500 equal weighted index, 20-year Strips  20-year US government zero coupon bonds.

10.5) 14.8) 0.45 44.7) 0.24

1994 /2005 Return Std Dev Sharpe R. M. drawdown Ret/drawdown

S&P 500

TABLE 6.2 Summary Statistics for Asset Categories

114 j CHAPTER 6

IMPROVING PERFORMANCE FOR LONG-TERM INVESTORS

j

115

has led to increasing interest in hedge funds. Many experts believe that the median future returns for hedge funds are likely to be lower than historical values—due in part to the large number of managers entering the arena. And as we will see, in fact low volatility may be a detriment for increasing overall portfolio performance by means of excess growth. Similarly, there are advantages to combining assets with inferior Sharpe ratios and reasonable returns, when these inferior values are caused by higher volatility. Interestingly, by comparing the capital-weighted S&P 500 index with an equal-weighted S&P 500,4 we see that the returns for the latter are higher than the former, and with higher volatility. The capital weighed index approximates a buy-and-hold portfolio. Extra volatility improves overall portfolio growth for the equal weighted index (as expected from the theoretical results of the previous section). To a degree, the equal-weighted index achieves rebalancing gains, but also displays a tilt to mid-size over the largest companies in the S&P 500 index. A similar issue pertains to the long-government bond index versus the 20-year strip index. The strip category has a lower Sharpe ratio (0.43 versus 0.66) due to the extra volatility embedded in the index. Strips are penalized by higher volatility. In contrast, Table 6.1 depicts the superior performance of the equal-weighted equity/strip portfolios over the traditional equity/bond portfolios. For the modified portfolio, not only is the Sharpe ratio higher for the fixed-mix 70/30 portfolio, but the excess returns are higher due to the higher volatility of the portfolio components—12.5) geometric return versus 11.9) for the static portfolio. The modified 70/30 mix, although much better than the traditional 70/30 mix, moderately under performs during 2000 2005 as compared with the earlier period—10.6) versus 14.4), respectively. The modified 60/40 mix performs better over the second time period, for a slightly more robust result due, in part, to the rebalancing gains obtained from the fixed-mix rule. What else can be done to increase performance vis-a`-vis the 70/30 and 60/40 benchmarks? As a first idea, we might try adding leverage to the benchmarks.5 While the returns increase with leverage (Table 6.3), the two risk measures also increase so that risk adjusted returns remains modest—Sharpe ratios around 0.55 and return/drawdown around 0.30. Increasing leverage does not improve the situation. The large drawdown values persist during the 2000 2005 period. As the next idea, we strive to achieve much wider diversification among the asset categories in our portfolio. To this end, we assemble an equally weighted mix (10) each) across 10 asset categories. The resulting fixed-mix portfolio takes a neutral view of any particular asset category, except that we disfavour assets with ultra low volatility (t-bills). The resulting portfolio displays much better performance over the full 12-year period and the two sub-periods (Table 6.4). In particular, the widely diversified portfolio can be levered to achieve 10) to 15) returns with reasonable drawdowns (under 15)). The risk adjusted returns are much better than the previous benchmarks (with or without leverage). Clearly, there are advantages to wide diversification and leverage in a fixed-mix portfolio. /

/

4 5

Rydex Investments sponsors an exchange traded fund with equal weights on the S&P 500 index. See Mulvey (2005). We charge t-bill rates here for leverage. Most investors will be required to pay additional fees.

116 j CHAPTER 6 TABLE 6.3 Historical Results of Leverage Applied to the Fixed-Mix 70/30 Asset Mix (Higher Returns Are Possible, but with Higher Volatility)

1994 /2005

1994 /1999

2000 /2005

Leverage S&P 500 equal weighted index LB Agg. Geometric return Standard deviation Sharpe ration Maximum drawdown Return/drawdown Geometric return Standard deviation Sharpe ration Maximum drawdown Return/drawdown Geometric return Standard deviation Sharpe ration Maximum drawdown Return/drawdown

0) 70) 30) 9.7) 10.4) 0.56 28.9) 0.33 18.2) 10.0) 1.32 10.2) 1.77 1.8) 10.4) 0.09 28.9) 0.06

20) 70) 30) 10.7) 12.5) 0.55 34.8) 0.31 20.9) 12.0) 1.33 12.4) 1.68 1.4) 12.5) 0.10 34.8) 0.04

50) 70) 30) 12.2) 15.6) 0.54 42.9) 0.29 24.9) 15.0) 1.33 15.7) 1.59 0.9) 15.7) 0.12 42.9) 0.02

100) 70) 30) 14.6) 20.8) 0.52 54.6) 0.27 31.7) 20.0) 1.34 21.1) 1.50 -0.3) 21.0) 0.15 54.6) 0.01

TABLE 6.4 Historical Results of Leverage Applied to a Widely Diversified Fixed-Mix Asset Mix (Each Asset Takes 10) Allocation—Excellent Risk-Adjusted Performance)

1994 /2005

1994 /1999

2000 /2005

Leverage LB Agg. EAFE NAREIT GSCI Hedge fund index CSFB managed futures index Currency index Tremont long/short S&P 500 equal weighted index LB 20-year STRIPS index Geometric return Standard deviation Sharpe ration Maximum drawdown Return/drawdown Geometric return Standard deviation Sharpe ration Maximum drawdown Return/drawdown Geometric return Standard deviation Sharpe ration Maximum drawdown Return/drawdown

0) 10) 10) 10) 10) 10) 10) 10) 10) 10) 10) 9.8) 6.2) 0.96 6.4) 1.54 9.6) 6.3) 0.75 6.4) 1.51 9.9) 6.2) 1.16 4.7) 2.10

20) 10) 10) 10) 10) 10) 10) 10) 10) 10) 10) 10.9) 7.4) 0.96 8.0) 1.37 10.5) 7.5) 0.75 8.0) 1.32 11.3) 7.4) 1.16 6.3) 1.81

50) 10) 10) 10) 10) 10) 10) 10) 10) 10) 10) 12.7) 9.3) 0.95 10.4) 1.22 11.9) 9.4) 0.74 10.4) 1.14 13.5) 9.3) 1.16 8.7) 1.56

100) 10) 10) 10) 10) 10) 10) 10) 10) 10) 10) 15.6) 12.4) 0.95 14.4) 1.08 14.1) 12.5) 0.73 14.4) 0.98 17.2) 12.4) 1.16 12.6) 1.35

IMPROVING PERFORMANCE FOR LONG-TERM INVESTORS

j

117

As the final improvement, we apply three overlay variables (commodities, currencies and fixed income) as defined in Section 6.1. These variables employ trend following rules based on the longstanding Mt. Lucas Management (MLM) index. Mulvey et al. (2004) evaluate the index with regard to re-balancing gains and related measures. The MLM index has produced equity-like returns with differential patterns over the past 30 years. However, an important change is made for our analysis. Rather than designating t-bills for the margin capital requirements, we assign the core assets (xi variables) for the margin capital, consistent with the model defined in Section 6.1 (as would be the case for a multistrategy hedge fund). Table 6.5 displays the results. Here, we lever the overlay variables at three values—20), 50) and 100)—within a fixed-mix rule. In all cases, the overlays greatly improve the risk-adjusted returns over the historical period (Sharpe and return to drawdown ratios greater than 1 and 1.5, respectively). (See also Brennan and Schwartz 1998 where continuous dynamic programming is used to optimize a traditional three asset model with a Treasury bond futures overlay.) The performance is positively affected by the relatively high volatility of individual asset categories, increasing portfolio returns via rebalancing gains. The overlay variables with the fixed-mix rule markedly improved performance over the historical period. TABLE 6.5 Historical Results of Overlay Variables—Fixed-Mix (Overlays Are More Efficient than Simple Leverage)

1994 /2005

1994 /1999

2000 /2005

LB Agg. EAFE NAREIT GSCI Hedge fund index CSFB managed futures index Currency index Tremont long/short S&P 500 equal weighted index LB 20-year STRIPS index Mt. Lucas commodity index Mt. Lucas currency index Mt. Lucas fixed income index Geometric return Standard deviation Sharpe ration Maximum drawdown Return/drawdown Geometric return Standard deviation Sharpe ration Maximum drawdown Return/drawdown Geometric return Standard deviation Sharpe ration Maximum drawdown Return/drawdown

10) 10) 10) 10) 10) 10) 10) 10) 10) 10) 0) 0) 0) 9.8) 6.2) 0.96 6.4) 1.54 9.6) 6.3) 0.75 6.4) 1.51 9.9) 6.2) 1.16 4.7) 2.10

10) 10) 10) 10) 10) 10) 10) 10) 10) 10) 20) 20) 20) 12.4) 7.1) 1.21 5.8) 2.13 12.6) 6.7) 1.14 5.8) 2.15 12.3) 7.6) 1.26 5.4) 2.27

10) 10) 10) 10) 10) 10) 10) 10) 10) 10) 50) 50) 50) 16.5) 9.7) 1.30 9.2) 1.80 17.0) 8.6) 1.40 9.2) 1.85 15.9) 10.8) 1.23 8.0) 2.00

10) 10) 10) 10) 10) 10) 10) 10) 10) 10) 100) 100) 100) 23.0) 15.3) 1.26 15.0) 1.53 24.3) 13.5) 1.43 14.5) 1.68 21.7) 16.9) 1.12 15.0) 1.45

118 j CHAPTER 6

To summarize, the historical tests illustrate that (6.1) re-balancing gains are possible with assets displaying relatively high volatility within fixed-mix portfolios and that (6.2) including overlay variables and leverage via fixed-mix can result in excellent risk-adjusted performance (almost hedge fund acceptable—23) annual geometric returns). Herein to reduce data mining concerns, we did not change the asset proportions during the period, except to re-balance back to the target mix each month. Also, we did not optimize asset proportions on the historical data, again to minimize data mining. The next section takes up the advantages of applying overlays in multi-period (dynamic mix) optimization models.

6.4

A STOCHASTIC PROGRAMMING PLANNING MODEL

A stochastic program (SP) gives the investor greater opportunities to improve performance as a function of changing economic conditions. These models can be constructed in two basic ways: (1) asset only, or (2) asset and liability management. We focus on asset-only problems in this report.6 It is generally agreed that the equity risk premium changes over longer time periods. In response, a number of researchers have developed equity valuation models. Bakshi and Chen (2005) designed a ‘fair’ equity valuation model based on three correlated stochastic processes: interest rates, projected earnings growth and actual earnings. The parameters of these processes are calibrated with market data (mostly historical prices of assets). They showed that future prices of equity assets revert on average to the calculated fair values. This type of analysis can be applied directly to a financial planning model based on a stochastic program. We highlight here only the major features of a stochastic program. The appendix provides further details. Also, see Mulvey and Thorlacius (1998) and Mulvey et al. (2000). In multistage stochastic programs, the evolution of future uncertainties is depicted in terms of a scenario tree. Constructing such a tree requires attention to three critical issues: (1) the realism of the model equations, (2) calibration of the parameters and (3) procedures to extract the sample set of scenarios. The projection system should be evaluated with historical data (back-testing), as well as on an ongoing basis. Our model employs a scenario generator that has been implemented widely for pension plans and insurance companies—the CAP:Link system (Mulvey et al. 2000). The system develops a close connection between the government spot rate and other economic and monetary factors such as GDP, inflation and currency movements. These connections are described in a series of references including Mulvey and Thorlacius (1998) and Mulvey et al. (1999, 2000). To illustrate, we describe a pair of linked stochastic processes for modelling the long and short interest rates. We assume that the rates link together through a correlated white noise term and by means of a stabilizing term that keeps the long short spread under control. The resulting spot rates follow /

6

See Mulvey et al. (2000) for details of related issues in ALM.

IMPROVING PERFORMANCE FOR LONG-TERM INVESTORS

j

119

drt ¼ kr ð r  rt Þdt þ gr ðst  sÞdt þ fr rt1=2 dzr ; dl ¼ k ðl l Þdt þ g ðs  sÞdt þ f l 1=2 dz ; t

l

t

l

t

l t

l

st ¼ rt  lt ; where rt and lt are the short and long interest rate, respectively, st is the spread between long and short interest rates, and white noise terms dzr and dzl are correlated. The model parameters include: kr ; kl gr ; gl fr ; f l r l s

drift on short and long interest rates, drift on the spread between long and short interest rates, instantaneous volatility, mean reversion level of short rate, mean reversion level of long rate, mean reversion level of the spread between long and short interest rates.

The second step involves parameter estimation. We calibrate parameters as a function of a set of common factors over the multi-period horizon. For example, equity returns and bond returns link to interest rate changes and underlying economic factors. Mulvey et al. (1999) described an integrated parameter estimation approach. Also, Bakshi and Chen (2005) and Chen and Dong (2001) discuss a related approach based on market prices of assets. Given the scenarios of traditional asset returns derived by means of the CAP:Link system, we obtain return scenarios of the overlay variables by assuming that they are conditionally normally distributed with the traditional asset categories according to historical relationships. This assumption is often employed in scenario generators as a form of a mixture model;7 see Chatfield and Collins (1980) for a discussion of the general properties. For our tests, we developed a condensed stochastic program in order to illustrate the issues for long-term investors. To this end, the resulting scenario tree is defined over a nine-year planning period, with three three-year time steps. The resulting problem consists of a modest (by current standards) nonlinear program. In this chapter, we employ a tree with 500 scenarios.8 The corresponding stochastic program contains 22,507 decision variables and 22,003 linear constraints. On average, it takes 10 20 s to solve for each point on the efficient frontier using a PC. Much larger stochastic programs are readily solvable with modern computers. See Dempster and Consigli (1998), Dempster et al. (2003) and Zenios and Ziemba (2006) for examples. Assume that an investor can invest her capital in the following core asset categories: treasury bills, S&P 500 index and 20-year zero coupon government bonds (STRIPS). /

7

The characteristics of the return series for the overlay strategies will be similar to those of the independent variables, such as mean reversion of interest rates and bond returns or fat tailed distributions. 8 There are 10 arcs (states) pointing out of the first node, followed by 50 arcs pointing out of the second stage nodes.

120 j CHAPTER 6 TABLE 6.6 Summary Statistics of Scenario Data for Six Asset Classes across a Nine-Year Planning Period T-bills Expected return Standard deviation Sharpe ratio Correlation matrix T-bills S&P 500 20-year STRIPS Commodity index Currency index Fixed income index

3.17) 0.016 0.000 1.000 0.017 0.030 0.004 0.009 0.009

S&P 500

20-year STRIPS

9.58) 0.172 0.374

8.13) 0.284 0.175

0.017 1.000 0.008 0.165 0.110 0.284

0.030 0.008 1.000 0.003 0.024 0.016

Commodity index 6.49) 0.106 0.313 0.004 0.165 0.003 1.000 0.002 0.028

Currency index 3.29) 0.058 0.021 0.009 0.110 0.024 0.002 1.000 0.164

Fixed income index 1.18) 0.057 0.347 0.009 0.284 0.016 0.028 0.164 1.000

In addition, she has the option to add the three previous overlay variables—commodity, currency and fixed income futures indices. As mentioned, in the scenario generator, we applied a conditional mean/covariance approach, based on historical relations between the core assets and the overlay variables. Remember that the overlays do not require any capital outlay. Table 6.6 lists the summary statistics of the generated scenario data for the six asset classes. Next, under a multi-period framework, we can calculate sample efficient frontiers, with any two objectives, for example, portfolio expected geometric returns and portfolio volatility9—Z1 and Z2 (Appendix). The first stochastic program forces the investor to invest solely in traditional assets. In contrast, under the latter two stochastic programs, the investor is allowed to add the overlays up to a given bound: 200) of the wealth at any time period in the second case and 300) in the last case. Note that unlike the historical back tests (previous sections), the allocation among overlays is no longer a fixed-mix, but is determined by the recommendation of the optimization and therefore will vary dynamically across time periods and scenarios. Figure 6.1 displays the illustrative efficient frontier under the described investment constraints.10 As expected, the solutions possessing the higher overlay bound dominate those under the lower overlay bounds and, as before, the overlay results dominate the traditional strategy. The larger the overlay bound, the greater potential to obtain higher returns at the cost of higher volatilities. For each efficient frontier, Table 6.7 lists descriptive statistics and asset allocations for the first period of three selected points on each frontier: the maximum return point, the minimum risk point and a compromise solution. Several observations are noteworthy. First, for all points under both strategies, the optimization chooses to invest either exclusively or dominantly in the commodity index for the first period. Second, although 9 We advocate that the investor evaluate a wide range of risk measures. These two are employed for illustration purposes. Mulvey et al. (2007) discuss real-world, multi-objective issues. 10 This model is a highly simplified version of an ALM system that has been implemented for the U.S. Department of Labor. The goals of the model are to assist pension plans in recovering their lost surpluses by optimizing assets in conjunction with managing liabilities (Mulvey et al. 2007). The unabridged system takes on the multi-objective environment discussed in the Appendix.

IMPROVING PERFORMANCE FOR LONG-TERM INVESTORS

j

121

Efficient Frontier 20% 18%

Annualized Return

16% 14% 12% 10% 8% 6% 4% 2% 0% 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 Risk (Std. deviation of return) Without overlays

200% Overlay bound

300% Overlay bound

FIGURE 6.1 Efficient frontiers under three investment constraints.

the minimum risk points display the smallest volatility, they tend to have large expected maximum drawdown compared with other points on the efficient frontier. For example, for the minimum point under 300) overlay bound, the volatility is 2) while the expected maximum drawdown is 23), and for the maximum point, although the volatility is 56), the expected maximum drawdown is only 6). Investors should choose a compromise point among the objectives that fits their risk appetite. In many cases, a compromise tends to possess reasonable trade-offs among the risk measures. As with the historical back tests, the three overlay assets improve investment performance. There are strong advantages to a stochastic program for assisting in financial planning. However, in most cases, a stochastic program is more complex to implement than a simple decision rule such as fixed-mix. Mulvey et al. (2007) demonstrated that running a stylized stochastic program can be helpful in discovering novel, improved policy rules. The results of the stochastic program show that the overlay (trend following approach) for commodities provides the best marginal risk/reward characteristics as compared with the other overlays (trend following for currencies and fixed income). In practice, alternative investment strategies should be considered. For example, Crownover (2006) shows that a combination of strategies (combined with a z-score approach) improves performance for currencies. These and related concepts can be readily applied via the discussed models.

6.5

SUMMARY AND FUTURE DIRECTIONS

This chapter shows that risk-adjusted performance can be enhanced by adding specialized overlays to multi-stage portfolios. We improved the returns in both the historical backtests with the fixed-mix rule and the stochastic programs. In the former case, with wide

Statistics Exp. annualized geo. ret. Exp. annualized ret. Standard deviation Mean (max drawdown) Stdev (max drawdown) Period 1 Alloc. T-bills S&P 500 20-year STRIPS Commodity index Currency index Fixed income index

Point on efficient frontier 7.78) 8.12) 0.054 0.000 0.000 0.00) 76.98) 23.02) 0.00) 0.00) 0.00)

0.00) 100.00) 0.00) 0.00) 0.00) 0.00)

Compromise

9.16) 10.05) 0.221 0.028 0.060

Max.

Without overlays

100.00) 0.00) 0.00) 0.00) 0.00) 0.00)

3.11) 3.17) 0.017 0.000 0.000

Min.

TABLE 6.7 Statistics for Selected Points on the Efficient Frontiers

0.00) 100.00) 0.00) 200.00) 0.00) 0.00)

13.19) 14.70) 0.351 0.035 0.092

Max.

0.00) 100.00) 0.00) 200.00) 0.00) 0.00)

11.10) 11.84) 0.117 0.016 0.050

Compromise

200) overlay bound

100.00) 0.00) 0.00) 199.63) 0.20) 0.09)

3.27) 3.33) 0.016 0.137 0.067

Min.

0.00) 100.00) 0.00) 300.00) 0.00) 0.00)

15.25) 17.51) 0.555 0.055 0.133

Max.

0.00) 100.00) 0.00) 300.00) 0.00) 0.00)

13.04) 14.17) 0.208 0.038 0.087

Compromise

300) overlay bound

100.00) 0.00) 0.00) 268.69) 16.99) 9.54)

3.27) 3.33) 0.016 0.231 0.059

Min.

122 j CHAPTER 6

IMPROVING PERFORMANCE FOR LONG-TERM INVESTORS

j

123

diversification, the investor benefits by the overlays within a fixed-mix rule. In this context, we demonstrated that long-term investors can take advantage of the overlay’s relatively high volatility. As mentioned, improvements are most appropriate for taxadvantaged investors such as pension trusts. Further, the re-balancing gains are available at the fund level (e.g. equal weighted S&P 500), and at the asset allocation level (for fixedmix). The use of overlays provides significant opportunities when linked with core asset categories requiring capital outlays: (1) there are potential roll returns11 via the futures markets—which reduce leverage costs, and (2) transaction costs can be often minimized through the liquid futures market. Importantly, the Sharpe ratios, while helpful at the portfolio level, can be misleading regarding the marginal benefit of an asset category within a fixed-mix portfolio context since volatility is penalized. What are potential implementation barriers? First, the fixed-mix rule requires rebalancing the asset mix at the end of each time period. The investor must take a disciplined approach, even in the face of large swings in asset returns, and be able to invest in a wide range of assets. Also, transaction costs must be considered; Mulvey and Simsek (2002) discuss approaches for addressing transaction costs through no-trade zones. Another possible barrier involves institutional legal constraints. Additionally individual investors may be unable to deploy equity and related assets as margin capital for their futures positions.12 In the domain of stochastic programs, we saw that the overlays can be beneficial as well. However, the resulting model grows exponentially as a function of the number of time periods and scenarios. For our simple example, the nine-year model with three-year time periods allows minimal re-balancing. A more realistic stochastic program with a greater number of time periods, while larger, would improve the trade-off between re-balancing gains and the extra returns derived from dynamic asset allocation in the face of changing economic conditions. Current computational power and greater information regarding the economic environment and patterns of asset prices helps overcome this barrier. A continuing research topic involves the search for assets with novel patterns of return (driven by factors outside the usual triple—interest rates, earnings and the general level of risk premium). An example might involve selling a limited amount of catastrophe insurance for hurricanes and earthquakes, or perhaps, taking on other weather related risks. Another example would be the numerous long short equity strategies that have become available. While these securities/strategies are not currently treated as asset categories, numerous novel futures-market instruments are under development; several have been recently implemented by exchanges such as the CME and CBOT. Undoubtedly, some of these instruments will help investors improve their risk-adjusted performance by achieving wider diversification in conjunction with selective leverage. /

11

Positive roll returns are possible when the futures market is in contango and the investor has a short position, or when backwardation occurs and the investor has a long position in the futures market. 12 The decision depends upon the arrangement with the investor’s prime broker. Swaps are ideal in this regard.

124 j CHAPTER 6

REFERENCES Bakshi, G.S. and Chen, Z., Stock valuation in dynamic economics. J. Financ. Mkts, 2005, 8, 115/151. Bell, D., Risk, return, and utility. Mgmt. Sci, 1995, 41, 23 /30. Brennan, M.J. and Schwartz, E.S., The use of treasury bill futures in strategic asset allocation programs. In Worldwide Asset and Liability Modeling, edited by W.T. Ziemba and J.M. Mulvey, pp. 205–228, 1998 (Cambridge University Press: Cambridge). Chatfield, C. and Collins, A.J., Introduction to Multivariate Analysis, 1980 (Chapman & Hall: New York). Chen, Z. and Dong, M., Stock valuation and investment strategies. Yale ICF Working Paper No. 00-46, 2001. Crownover, C., Dynamic capital allocation: Exploiting persistent patterns in currency performance. Quant. Finance, 2006, 6(3), 185/191. Dempster, M.A.H. and Consigli, G., The CALM stochastic programming model for dynamic assetliability management. In Worldwide Asset and Liability Modeling, edited by W.T. Ziemba and J.M. Mulvey, pp. 464–500, 1998 (Cambridge University Press: Cambridge). Dempster, M.A.H., Evstigneev, I.V. and Schenk-Hoppe´ , K.R., Volatility-induced financial growth. Quant. Finance, April 2007. Dempster, M.A.H., Germano, M., Medova, E.A. and Villaverde, M., Global asset liability management. Br. Actuarial J., 2003, 9, 137/216. Fernholz, R., Stochastic Portfolio Theory, 2002 (Springer-Verlag: New York). Fernholz, R. and Shay, B., Stochastic portfolio theory and stock market equilibrium. J. Finance, 1982, 37, 615/624. Luenberger, D., Investment Science, 1997 (Oxford University Press: New York). Merton, R.C., Lifetime portfolio selection under uncertainty: The continuous-time case. Rev. Econ. Stat., 1969, 51, 247/257. Mulvey, J.M., Essential portfolio theory. A Rydex Investment White Paper, 2005, (also Princeton University report). Mulvey, J.M., Gould, G. and Morgan, C., An asset and liability management system for Towers Perrin-Tillinghast. Interfaces, 2000, 30, 96 /114. Mulvey, J.M., Kaul, S.S.N. and Simsek, K.D., Evaluating a trend-following commodity index for multi-period asset allocation. J. Alternative Invest., 2004, Summer, 54 /69. Mulvey, J.M., Shetty, B. and Rosenbaum, D., Parameter estimation in stochastic scenario generation systems. Eur. J. Operat. Res., 1999, 118, 563/577. Mulvey, J.M. and Simsek, K.D., Rebalancing strategies for long-term investors. In Computational Methods in Decision-Making, Economics and Finance: Optimization Models, edited by E.J. Kontoghiorghes, B. Rustem and S. Siokos, pp. 15–33, 2002 (Kluwer Academic Publishers: Netherlands). Mulvey, J.M., Simsek, K.D., Zhang, Z., Fabozzi, F. and Pauling, W., Assisting defined-benefit pension plans. Operat. Res., 2008. Mulvey, J.M. and Thorlacius, A.E., The Towers Perrin global capital market scenario generation system: CAP Link. In Worldwide Asset and Liability Modeling, edited by W.T. Ziemba and J.M. Mulvey, 1998 (Cambridge University Press: Cambridge). Samuelson, P.A., Lifetime portfolio selection by dynamic stochastic programming. Rev. Econ. Stat., 1969, 51, 239/246. Zenios, S.A. and Ziemba, W.T., (Eds), Handbook of Asset and Liability Modeling, 2006 (NorthHolland: Amsterdam).

IMPROVING PERFORMANCE FOR LONG-TERM INVESTORS

j

125

APPENDIX 6.A: MATHEMATICAL MODEL FOR THE MULTI-STAGE STOCHASTIC PROGRAM This appendix defines the asset-only investment problem as a multi-stage stochastic program. We define the set of planning periods as T ¼ f0; 1; . . . ; t; t þ 1g. We focus on the investor’s position at the beginning of period t  1. Decisions occur at the beginning of each time period. Under a multi-period framework, we assume that the portfolio is rebalanced at the beginning of each period. Asset investment categories—assets requiring capital outlays—are defined by set I ¼ f1; 2; . . . ; N g, with category 1 representing cash. The remaining categories can include broad investment groupings such as growth and value domestic and international stocks, long-term government and corporate bonds, and real estate. The overlay variables are defined by set J ¼ f1; 2; . . . ; Mg. Uncertainty is represented by a set of scenarios s  S. The scenarios may reveal identical values for the uncertain quantities up to a certain period—i.e. they share common information history up to that period. We address the representation of the information structure through non-anticipativity constraints, which require that variables sharing a common history, up to time period t, must be set equal to each other. For each i  I, j  J, t T and s  S, we define the following parameters and decision variables: Parameters ri;t;s ¼ 1 þ ri;t;s ; where ri;t;s is return of traditional asset i in period t, under scenario s (e.g. Mulvey et al. (2000)). rj;t;s ¼ 1 þ rj;t;s , where rj;t;s is return of overlay asset j in period t, under scenario s. P Probability that scenario s occurs - s2S ps ¼ 1: ps 7! xi;0;s Amount allocated to traditional asset class i, at the end of period 0, under scenario s, before first rebalancing. si;t Transaction costs for rebalancing asset i in period t (symmetric transaction costs are assumed). OL B Total overlay bound. G TA Target assets at the horizon. Decision variables xi;t;s 7! xi;t;s BUY xi;t;s SELL xi;t;s

Amount allocated to traditional asset class i, at the beginning of period t, under scenario s, after rebalancing. Amount allocated to traditional asset class i at the end of period t, under scenario s, before rebalancing. Amount of traditional asset class i purchased for rebalancing in period t, under scenario s. Amount of traditional asset class i sold for rebalancing in period t, under scenario s.

126 j CHAPTER 6

xj;t;s 7! xj;t;s

Amount allocated to overlay asset j, at the beginning of period t, under scenario s. Amount allocated to overlay asset j at the end of period t, under scenario s.

TA7! xt;s

Asset wealth at the end of time period t, under scenario s.

Given these definitions, we present the deterministic equivalent of the stochastic asset-only allocation problem: Model [MSP] (general structure) U fZ1 ; Z2 ; . . . ; Zk g;

Maximize

ð6:A1Þ

where the goals are defined as functions of the decision variables (examples of various goals are shown below): Zk ¼ fk ðxÞ; subject to: X

7! TA7! xi;0;s ¼ x0;s

8 s 2 S;

ð6:A2Þ

i2I

X

7! xi;t;s þ

i2I

X

7! TA7! xj;t;s ¼ xt;s

8 s 2 S; t ¼ 1; . . . ; t þ 1;

i2J

7! xi;t;s ¼ ri;t;s xi;t;s

8 s 2 S; t ¼ 1; . . . ; t; i 2 I;

7! ¼ ðrj;t;s  1Þxj;t;s xj;t;s

X

8 s 2 S; t ¼ 1; . . . ; t; j 2 J ;

! xj;t;s B OL xtTA7 1;s

8 s 2 S; t ¼ 1; . . . ; t

j2J

7! BUY SELL xi;t;s ¼ xi;t1;s þ xi;t1;s ð1  si;t1 Þ  xi;t1;s

7! x1;t;s ¼ x1;t1;s þ

X

SELL xi;t1;s ð1  si;t1 Þ 

i6¼1

 bt1;s þ

8 s 2 S; i 6¼ 1; t ¼ 1; . . . ; t þ 1;

CONT yt1;s

X

ð6:A3Þ

ð6:A4Þ

ð6:A5Þ

ð6:A6Þ

ð6:A7Þ

BUY xi;t1;s

i6¼1

8 s 2 S; t ¼ 1; . . . ; t þ 1;

ð6:A8Þ

IMPROVING PERFORMANCE FOR LONG-TERM INVESTORS

xi;t;s1 ¼ xi;t;s2 ;xj;t;s1 ¼ xj;t;s2

j

127

ð6:A9Þ

8 s1 and s2 with identical past up to time t;

RiskfZ1 ; Z2 ; . . . ; Zk g Riskmax :

ð6:A10Þ

The objective function (6.A1) depicts a generic multi-objective optimization problem. It can take several forms. For instance, we could employ the von Neumann Morgenstern utility function of the final wealth. Alternatively, we could use the classical return-risk function Z ¼ Z MeanðxtTA7! Þþ ð1  ZÞ RiskðxtTA7! Þ, where MeanðxtTA7! Þ is the expected total assets at the end of t and RiskðxtTA7! Þ is a risk measure of the total final wealth across all scenarios. The weight parameter h indicates the relative importance of risk as compared with expected wealth. Constraint (6.A2) represents the initial total value of assets at the end of period 0. Constraint (6.A3) depicts wealth at the end of period t, aggregating assets in traditional asset classes and investment gains/losses from overlay strategies. The wealth accumulated at the end of period t before rebalancing in traditional asset class i is given by (6.A4). The wealth accumulation due to overlay variable j at the end of period t is depicted in (6.A5). Constraint (6.A6) sets the bound for overlays for each time period and across all scenarios. The flow balance constraints for all traditional asset classes except cash, for all periods, are given by (6.A7). (6.A8) represents the flow balance constraint for cash. Non-anticipativity constraints are represented by (6.A9), ensuring that the scenarios with the same past will have identical decisions up to that period. Risk-based constraints appear in (6.A10). Here we list a few popular goals among numerous others. Especially, we set GTA to be the target wealth for the investor at t1. The first goal is to maximize the expected final investor wealth at the horizon: /

Z1 ¼

X

TA7! ps xt;s :

s

Both the second and the third goals quantify the risk of missing the target wealth at the planning horizon. Goal 2 is the downside risk of the expected final wealth: þ 2 X  TA7! ps Z1  xt;s : Z2 ¼ s

A similar goal is the downside risk of the expected final investor wealth with respect to target wealth GTA at the horizon: þ 2 X  TA TA7! Z3 ¼ ps G  xt;s : s

Goal 3 is zero if and only if final wealth reaches the target under all scenarios.

128 j CHAPTER 6

The fourth and fifth goals focus on the timing of achieving the target wealth. Goal 4 measures the expected earliest time for the investor’s assets to reach the target: Z4 ¼

X

n o TA7! ps inf t : xt;s  G TA :

s

The goal Z4 could be greater than t1 if the wealth could not reach the target before the planning horizon. Aiming at measuring the risk of missing the target date of reaching the goal, we propose the fifth goal, the downside risk of the time to achieve the goal: o þ 2 X  n TA7! TA t ps inf t : xt;s  G : Z5 ¼ s

The model could be readily modified to incorporate liability-related decisions and other investment strategies. For instance, the fixed-mix rule enforces the following constraint at each juncture: li ¼

xi;t;s TA xt;s

;

for any time period t and under any scenario s;

where xt;s is the total wealth at the beginning of period t and we define the proportion of wealth to be li for each asset i I. Ideally, we would maintain the target proportion l at all time periods and under every scenario. Rebalancing under a fixed-mix rule automatically ‘buys low and sells high.’ However, the fixed-mix constraints induce non-convexity into the stochastic program. Specially designed algorithms are needed to solve such a problem.

CHAPTER

7

Stochastic Programming for Funding Mortgage Pools GERD INFANGER CONTENTS 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 7.2 Funding Mortgage Pools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 7.2.1 Interest Rate Term Structure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 7.2.2 The Cash Flows of a Mortgage Pool . . . . . . . . . . . . . . . . . . . . . . . 133 7.2.3 Funding through Issuing Debt . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 7.2.4 Leverage Ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 7.2.5 Simulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 7.2.6 The Net Present Value of the Payment Stream . . . . . . . . . . . . . . . . 136 7.2.7 Estimating the Expected NPV of the Payment Stream . . . . . . . . . . . 136 7.2.8 The Expected NPV of a Funding Mix . . . . . . . . . . . . . . . . . . . . . . 137 7.2.9 Risk of a Funding Mix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 7.3 Single-Stage Stochastic Programming . . . . . . . . . . . . . . . . . . . . . . . . . . 138 7.4 Multi-Stage Stochastic Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 7.5 Duration and Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 7.5.1 Effective Duration and Convexity . . . . . . . . . . . . . . . . . . . . . . . . . 143 7.5.2 Traditional Finance: Matching Duration and Convexity. . . . . . . . . . 145 7.5.3 Duration and Convexity in the Single-Stage Model . . . . . . . . . . . . 146 7.5.4 Duration and Convexity in the Multi-Stage Model . . . . . . . . . . . . . 146 7.6 Computational Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 7.6.1 Data Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 7.6.2 Results for the Single-Stage Model . . . . . . . . . . . . . . . . . . . . . . . . 149 7.6.3 Results for the Multi-Stage Model . . . . . . . . . . . . . . . . . . . . . . . . 151 7.6.4 Results for Duration and Convexity . . . . . . . . . . . . . . . . . . . . . . . 153 7.6.5 Duration and Convexity, Multi-Stage Model . . . . . . . . . . . . . . . . . 154 7.6.6 Out-of-Sample Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158 7.6.7 Larger Sample Size. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 7.7 Large-Scale Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163 129

130 j CHAPTER 7

7.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . Acknowledgements . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . Appendix 7.A: Tables and Graphs for Data Sets

7.1

H

.... .... .... ‘Flat’

......... ......... ......... and ‘Steep’ .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

165 166 166 168

INTRODUCTION ISTORICALLY, THE BUSINESS OF CONDUITS,

like Freddie Mac, Fannie Mae or Ginnie Mae, has been to purchase mortgages from primary lenders, pool these mortgages into mortgage pools, and securitize some if not all of the pools by selling the resulting Participation Certificates (PCs) to Wall Street. Conduits keep a fixed markup on the interest for their profit and roll over most of the (interest rate and prepayment) risk to the PC buyers. Recently, a more active approach, with the potential for significantly higher profits, has become increasingly attractive: instead of securitizing, funding the purchase of mortgage pools by issuing debt. The conduit firm raises the money for the mortgage purchases through a suitable combination of long- and short-term debt. Thereby, the conduit assumes a higher level of risk due to interest rate changes and prepayment risk but gains higher expected revenues due to the larger spread between the interest on debt and mortgage rates compared with the fixed markup by securitizing the pool. The problem faced by the conduits is an asset-liability management problem, where the assets are the mortgages bought from primary lenders and the liabilities are the bonds issued. Asset liability problems usually are faced by pension funds and insurance companies. Besides assets, pension plans need to consider retirement obligations, which may depend on uncertain economic and institutional variables, and insurance companies need to consider uncertain pay-out obligations due to unforseen and often catastrophic events. Asset liability models are most useful when both asset returns and liability payouts are driven by common, e.g. economic, factors. Often, the underlying stochastic processes and decision models are multi-dimensional and require multiple state variables for their representation. Using stochastic dynamic programming, based on Bellman’s (1957) dynamic programming principle, for solving such problems is therefore computationally difficult, well known as the ‘curse of dimensionality.’ If the number of state variables of the problem is small, stochastic dynamic programming can be applied efficiently. Infanger (2006) discusses a stochastic dynamic programming approach for determining optimal dynamic asset allocation strategies over an investment horizon with many re-balancing periods, where the value-to-go function is approximated via Monte Carlo sampling. The chapter uses an in-sample/out-of-sample approach to avoid optimization bias. Stochastic programming can take into account directly the joint stochastic processes of asset and liability cash flows. Traditional stochastic programming uses scenario trees to represent possible future events. The trees may be constructed by a variety of scenariogeneration techniques. The emphasis is on keeping the resulting tree thin but representative of the event distribution and on arriving at a computationally tractable problem, where obtaining a good first-stage solution rather than obtaining an entire accurate policy is the

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

131

goal. Early practical applications of stochastic programming for asset liability management are reported in Kusy and Ziemba (1986) for a bank and in Carino et al. (1994) for an insurance company. Ziemba (2003) gives a summary of the stochastic programming approach for asset liability and wealth management. Early applications of stochastic programming for asset allocation are discussed in Mulvey and Vladimirou (1992), formulating financial networks, and Golub et al. (1995). Examples of early applications of stochastic programming for dynamic fixed-income strategies are Zenios (1993), discussing the management of mortgage-backed securities, Hiller and Eckstein (1993), and Nielsen and Zenios (1996). Wallace and Ziemba (2005) present recent applications of stochastic programming, including financial applications. Frauendorfer and Schu¨ erle (2005) discuss the re-financing of mortgages in Switzerland. Monte Carlo sampling is an efficient approach for representing multi-dimensional distributions. An approach, referred to as decomposition and Monte Carlo sampling, uses Monte Carlo (importance) sampling within a decomposition for estimating Benders cut coefficients and right-hand sides. This approach has been developed by Dantzig and Glynn (1990) and Infanger (1992). The success of the sampling within the decomposition approach depends on the type of serial dependency of the stochastic parameter processes, determining whether or not cuts can be shared or adjusted between different scenarios of a stage. Infanger (1994) and Infanger and Morton (1996) show that, for serial correlation of stochastic parameters (in the form of autoregressive processes), unless the correlation is limited to the right-hand side of the (linear) program, cut sharing is at best difficult for more than three-stage problems. Monte Carlo pre-sampling uses Monte Carlo sampling to generate a tree, much like the scenario-generation methods referred to above, and then employs a suitable method for solving the sampled (and thus approximate) problem. We use Monte Carlo pre-sampling for representing the mortgage funding problem, and combine optimization and simulation techniques to obtain an accurate and tractable model. We also provide an efficient way to independently evaluate the solution strategy from solving the multi-stage stochastic program to obtain a valid upper bound on the objective. The pre-sampling approach provides a general framework of modeling and solving stochastic processes with serial dependency and many state variables; however, it is limited in the number of decision stages. Assuming a reasonable sample size for representing a decision tree, problems with up to four decision stages are meaningfully tractable. Dempster and Thorlacius (1998) discuss the stochastic simulation of economic variables and related asset returns. A recent review of scenario-generation methods for stochastic programming is given by Di Domenica et al. (2006), discussing also simulation for stochastic programming scenario generation. In this chapter we present how multi-stage stochastic programming can be used for determining the best funding of a pool of similar fixed-rate mortgages through issuing bonds, callable and non-callable, of various maturities. We show that significant profits can be obtained using multi-stage stochastic programming compared with using a singlestage model formulation and compared with using duration and convexity hedging, strategies often used in traditional finance. For the comparison we use an implementation

132 j CHAPTER 7

of Freddie Mac’s interest rate model and prepayment function. We describe in Section 7.2 the basic formulation of funding mortgage pools and discuss the estimation of expected net present value and risk for different funding instruments using Monte Carlo sampling techniques. In Section 7.3 we discuss the single-stage model. In Section 7.4 we present the multi-stage model. Section 7.5 discusses duration and convexity and delta and gamma hedging. In Section 7.6 we discuss numerical results using practical data obtained from Freddie Mac. We compare the efficient frontiers from the single-stage and multi-stage models, discuss the different funding strategies and compare them with delta and gamma hedged strategies, and evaluate the different strategies using out-of-sample simulations. In particular, Section 7.6.5 presents the details of the out-of-sample evaluation of the solution strategy obtained from solving a multi-stage stochastic program. Section 7.7 reports on the solution of very large models and gives model sizes and solution times. Finally, Section 7.8 summarizes the results of the chapter. While not explicitly discussed in this chapter, the problem of what fraction of the mortgage pool should be securitized, and what portion should be retained and funded through issuing debt can be addressed through a minor extension of the models presented. Funding decisions for a particular pool are not independent of all other pools already in the portfolio and those to be acquired in the future. The approach can of course be extended to address also the funding of a number of pools with different characteristics. While the chapter focuses on funding a pool of fixed-rate mortgages, the framework applies analogously to funding pools of adjustable-rate mortgages.

7.2

FUNDING MORTGAGE POOLS

7.2.1

Interest Rate Term Structure

Well-known interest rate term structure models in the literature are Vasicek (1977), Cox et al. (1985), Ho and Lee (1986), and Hull and White (1990), based on one factor, and Longstaff and Schwarz (1992) based on two factors. Observations of the distributions of future interest rates are obtained using an implementation of the interest rate model of Luytjes (1993) and its update according to the Freddie Mac document. The model reflects a stochastic process based on equilibrium theory using random shocks for short rate, spread (between the short rate and the tenyear rate) and inflation. We do not use the inflation part of the model and treat it as a two-factor model, where the short rate and the spread are used to define the yield curve. To generate a possible interest rate path we feed the model at each period with realizations of two standard normal random variables and obtain as output for each period a possible outcome of a yield curve of interest rates based on the particular realizations of the random shocks. Given a realization of the short rate and the spread, the new yield curve is constructed free of arbitrage for all calculated yield points. We denote as it(m), t
1,. . .,T, the random interest rate of a zero coupon bond of term m in period t.

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

7.2.2

j

133

The Cash Flows of a Mortgage Pool

We consider all payments of a pool of fixed-rate mortgages during its lifetime. Time periods t range from t
0,. . .,T, where T denotes the end of the horizon; e.g. T
360 reflects a horizon of 30 years considering monthly payments. We let Bt be the balance of the principal of the pool at the end of period t. The principal capital B0 is given to the homeowners at time period t
0 and is regained through payments bt and through prepayments at at periods t
1,. . .,T. The balance of the principal is updated periodically by Bt ¼ Bt1 ð1 þ k0 Þ  bt  at ;

t ¼ 1; . . . ; T :

The rate k0 is the contracted interest rate of the fixed-rate mortgage at time t
0. We define lt to be the payment factor at period t
1,. . .,T. The payment factor when multiplied by the mortgage balance yields the constant monthly payments necessary to pay off the loan over its remaining life, e.g. lt ¼ k0 =ð1  ð1 þ k0 ÞtT 1 Þ; thus, bt ¼ lt Bt1 : The payment factor lt depends on the interest rate k0. For fixed-rate mortgages the quantity k0, and thus the quantities lt, are known with certainty. However, prepayments at, at periods t
1,. . .,T, depend on future interest rates and are therefore random parameters. Prepayment models or functions represent the relationship between interest rates and prepayments. See, for example, Kang and Zenios (1992) for a detailed discussion of prepayment models and factors driving prepayments. In order to determine at we use an implementation of Freddie Mac’s prepayment function according to Lekkas and Luytjes (1993). Denoting the prepayment rates obtained from the prepayment function as gt , t
1,. . ., T, we compute the prepayments at in period t as at ¼ gt Bt1 :

7.2.3

Funding through Issuing Debt

We consider funding through issuing bonds, callable and non-callable, with various maturities. Let ‘ be a bond with maturity m‘, ‘ 2 L, where L denotes the set of bonds under consideration. Let f‘tt be the payment factor for period t, corresponding to a bond issued at period t, t 5t 5t m‘ :

134 j CHAPTER 7

8 þ1; if t  t ¼ 0; > > > < if 0 < t  t < m‘ ; f‘tt  ðit ðm‘ Þ þ st‘ Þ; > > > : ð1 þ it ðm‘ Þ þ st‘ Þ; if t  t ¼ m‘ ; where it (m‘ ) reflects the interest rate of a zero coupon bond with maturity ml, issued at period t, and st‘ denotes the spread between the zero coupon rate and the actual rate of bond l issued at t. The spread st‘ includes the spread of bullet bonds over zero coupon bonds (referred to as agency spread) and the spread of callable bonds over bullet bonds (referred to as agency call spread), and is computed according to the model specification given in the Freddie Mac document (Luytjes 1996). Let Mt‘ denote the balance of a bond ‘ at the time t it is issued. The finance payments resulting from bond ‘ are dt‘ ¼ f‘tt Mt‘ ;

t ¼ t; . . . ; t þ m‘ ;

from the time of issue (t) until the time it matures (t ml) or, if callable, it is called. We consider the balance of the bullet from the time of issue until the time of maturity as Mt‘ ¼ Mt‘ ;

7.2.4

t ¼ t; . . . ; t þ m‘ :

Leverage Ratio

Regulations require that, at any time t, t
0,. . .,T, equity is set aside against debt in an amount such that the ratio of the difference of all assets minus all liabilities to all assets is greater than or equal to a given value m. Let Et be the balance of an equity (cash) account associated with the funding. The equity constraint requires that Bt þ Et  Mt Bt þ Et

 m;

where the total asset balance is the sum of the mortgage balance and the equity balance, P Bt  Et , and Mt ¼ ‘ Mt‘ is the total liability balance. At time periods t
0,. . .,T, given the mortgage balance Bt , and the liability balance Mt , we compute the equity balance that fulfills the leverage ratio constraint with equality as Et ¼

Mt  Bt ð1  mÞ ; 1m

t ¼ 0; . . . ; T :

We assume that the equity account accrues interest according to the short rate it(short), the interest rate of a 3-month zero coupon bond. Thus, we have the following balance equation for the equity account:

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

135

Et ¼ Et1 ð1 þ it1 ðshortÞÞ þ et1 ;

where et are payments into the equity account (positive) or payments out of the equity account (negative). Using this equation we compute the payments et to and from the equity account necessary to maintain the equity balance Et computed for holding the leverage ratio m. 7.2.5

Simulation

Using the above specification we may perform a (Monte Carlo) simulation in order to obtain an observation of all cash flows resulting from the mortgage pool and from financing the pool through various bonds. In order to determine in advance how the funding is carried out, we need to specify certain decision rules defining what to do when a bond matures, when to call a callable bond, at what level to fund, and how to manage profits and losses. For the experiment we employed the following six rules. (i) Initial funding is obtained at the level of the initial balance of the mortgage pool, M0
B0. (ii) Since at time t
0, M0
B0, it follows that E0 ¼ ½m=ð1  mÞ B0 , an amount that we assume to be an endowed initial equity balance. (iii) When a bond matures, refunding is carried out using short-term debt (non-callable 3-month bullet bond) until the end of the planning horizon, each time at the level of the balance of the mortgage pool. (iv) Callable bonds are called according to the call rule specification in Freddie Mac’s document (Luytjes 1996). Upon calling, refunding is carried out using short-term debt until the end of the planning horizon, each time at the level of the balance of the mortgage pool. (v) The leverage ratio (ratio of the difference of all assets minus all liabilities to all assets) is m
0.025. (vi) At each time period t, after maintaining the leverage ratio, we consider a positive sum of all payments as profits and a negative sum as losses. According to the decision rules, when funding a mortgage pool using a single bond ‘, we assume at time t
0 that M0‘ ¼ B0 , i.e. that exactly the amount of the initial mortgage balance is funded using bond ‘. After bond matures refunding takes place using another ^ based on the interest bond (according to the decision rules, short-term debt, say, bond ‘), rate and the level of the mortgage balance at the time it is issued. If the initial bond ‘ is callable, it may be called, and then funding carried out through another bond (say, short^ Financing based on bond ‘^ is continued until the end of the planning term debt ‘). horizon, i.e. until T  t < m‘^, and no more bond is issued. Given the type of bond being used for refunding, and given an appropriate calling rule, all finance payments for the initial funding using bond ‘ and the subsequent refunding using bond ‘^ can be determined. We denote the finance payments accruing from the initial funding based on bond ‘ and its consequent refunding based on bond ‘^ as

136 j CHAPTER 7

dt‘ ;

t ¼ 1; . . . ; T :

Once the funding and the corresponding liability balance Mt‘ is determined, the required equity balance Et ¼ Et‘ and the payments et ¼ et‘ are computed. 7.2.6

The Net Present Value of the Payment Stream

Finally, we define as Pt‘ ¼ bt þ at þ dt‘  et‘ ;

t ¼ 1; . . . ; T ; P0 ¼ M0  B0 ¼ 0

the sum of all payments in period t, t
0,. . .,T, resulting from funding a pool of mortgages (initially) using bond ‘. Let It be the discount factor for period t, i.e.

It ¼

t Y ð1 þ ik ðshortÞÞ;

t ¼ 1; . . . ; T ; I0 ¼ 1;

k¼1

where we use the short rate at time t, it(short), for discounting. The net present value (NPV) of the payment stream is then calculated as

r‘ ¼

T X Pt‘ t¼0

It

:

So far, we consider all quantities that depend on interest rates as random parameters. In particular, Pt‘ is a random parameter, since bt , at , dt‘ , and et‘ are random parameters depending on random interest rates. Therefore, the net present value r‘ is a random parameter as well. In order to simplify the notation we do not label any specific outcomes of the random parameters. A particular run of the interest rate model requires 2T random outcomes of unit normal random shocks. We now label a particular path of the interest rates obtained from one run of the interest rate model and all corresponding quantities with v. In particular, we label a realization of the net present value based on a particular interest rate path as r‘o . 7.2.7

Estimating the Expected NPV of the Payment Stream

We use Monte Carlo sampling to estimate the expected value of the NPV of a payment stream. Under a crude Monte Carlo approach to the NPV estimation, we sample N paths vS, N
|S|, using different observations of the distributions of the 2T random parameters as input to the interest rate model, and we compute r‘o for each v S. Then, an estimate for the expected net present value (NPV) of the cash flow stream based on initial funding using bond ‘ is

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

r‘ ¼

j

137

1 X o r‘ : N o2S

We do not describe in this document how we use advanced variance reduction techniques (e.g. importance sampling) for the estimation of the expected net present value of a payment stream. We refer to Prindiville (1996) for how importance sampling could be applied. 7.2.8

The Expected NPV of a Funding Mix

Using simulation (as described above) we compute the net present value of the payment stream r‘o for each realization v S and each possible initial funding ‘  L. The net present value of a funding mix is given by the corresponding convex combination of the net present values of the components ‘  L, i.e. ro ¼

X

X

r‘o x‘ ;

‘2L

x‘ ¼ 1;

xl  0;

‘2L

where x‘ are non-negative weights summing to one. The expected net present value of a funding mix, r ¼

1 X o r ; N o2S

is also represented as the convex combination of the expected net present values of the components ‘  L, i.e. r ¼

X

X

r‘ x‘ ;

‘2L

7.2.9

x‘ ¼ 1;

x‘  0:

‘2L

Risk of a Funding Mix

In order to measure risk, we use as an appropriate asymmetric penalty function the negative part of the deviation of the NPV of a funding portfolio from a pre-specified target u, i.e. vo ¼

X

! r‘o x‘  u

;

‘2L

and consider risk as the expected value of vv, estimated as v ¼

1 X o v : N o2S

138 j CHAPTER 7

A detailed discussion of this particular risk measure is given in Infanger (1996). The efficient frontier with risk as the first lower partial moment is also referred to as the ‘put call efficient frontier;’ see, for example, Dembo and Mausser (2000).

7.3

SINGLE-STAGE STOCHASTIC PROGRAMMING

Having computed the NPVs r‘o for all initial funding options ‘  L and for all paths v S, we optimize the funding mix with respect to expected returns and risk by solving the (stochastic) linear program min s.t.

1 X o v ¼ v; N X r‘o x‘ þ v o  u;

o 2 S;

‘

X

r‘ x‘  r;

‘

X

x‘ ¼ 1;

x‘  0;

v o  0:

‘

The parameter r is a pre-specified value that the expected net present value of the r‘ g. Using the model we portfolio should exceed or be equal to. Clearly, r  rmax ¼ max‘ f max and successively reducing r until trace out an efficient frontier starting with r ¼ r r
0, each time solving the linear program to obtain the portfolio with the minimum risk v corresponding to a given value of r. The single-stage stochastic programming model optimizes funding strategies based on decision rules defined over the entire planning horizon of T
360 periods, where the net present value of each funding strategy using initially bond ‘ and applying the decision rules is estimated using simulation. A variant of the model arises by trading off expected NPV and risk in the objective, with l denoting the risk-aversion coefficient: min s.t.



X

X

r‘ x‘ þ l

‘

1 X N

r‘o x‘ þ v o  u;

vo; o 2 S;

‘

X

x‘ ¼ 1;

x‘  0;

v o  0:

‘

For a risk aversion of l
0, risk is not part of the objective and expected NPV is maximized. The efficient frontier can be traced out by increasing the risk aversion l successively from zero to very large values, where the risk term in the objective entirely dominates. This approach is very different to Markowitz’s (1952) mean variance analysis in that the distribution of the NPV is represented through scenarios (obtained through simulations

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

139

over a long time horizon, considering the application of decision rules) and a downside risk measure is used for representing risk.

7.4

MULTI-STAGE STOCHASTIC PROGRAMMING

In the following we relax the application of decision rules at certain decision points within the planning horizon, and optimize the funding decisions at these points. This leads to a multi-stage stochastic programming formulation. We partition the planning horizon Ž0,T into n sub-horizons hT1 ; T2 i; hT2 ; T3 i; . . . ; hTn ; Tnþ1 i, where T1
0, and Tnþ1 ¼ T . For the experiment, we consider n
4, and partition at T1
0, T2
12, T3
60, and T4
360. We label the decision points at time t
T1 as stage 1, at time t
T2 as stage 2, and at time t
T3 as stage 3 decisions. Funding obtained at the decision stages is labeled as ‘1  L1, ‘2  L2, and ‘3  L3 according to the decision stages. At time t
T4, at the end of the planning horizon, the (stage 4) decision involves merely evaluating the net present value of each end point for calculating the expected NPV and risk. In between the explicit decision points, at which funding is subject to optimization, we apply the decision rules defined above. Instead of interest rate paths as used in the single-stage model, we now use an interest rate tree with nodes at each stage. We consider NS2N paths v2  S2 between t
T1 and t
T2; for each node v2  S2 we consider NS3N paths v3  S3 between t
T2 and t
T3; for each node ðo2 ; o3 Þ 2 fS2  S3 g we consider NS4N paths v4  S4 between t
T3 and t
T4. Thus, the tree has NS2S3S4N end points. We may denote S ¼ fS2  S3  S4 g and o ¼ ðo2 ; o3 ; o4 Þ. Thus a particular path through the tree is now labeled as o ¼ ðo2 ; o3 ; o4 Þ using an index for each partition. Figure 7.1 presents the decision tree of the multi-stage model for only two paths for each period. The simulation runs for each partition of the planning horizon are carried out in such a way that the dynamics of the interest rate process and the prepayment function are fully carried forward from one partition to the next. Since the interest rate model and the

ω2 S2

e.g.

< 1yr > T1

(ω2, ω3) (S2 × S3)

< 4yrs > T2

(ω2, ω3, ω4) (S2 × S3 × S4) = S

< 25yrs > T3

T4 Time

Now Decision node all decisions subject to optimization

FIGURE 7.1 Multi-stage model setup, decision tree.

Simulation path (monthly) application of decision rules

140 j CHAPTER 7

prepayment function include many lagged terms and require the storing of 64 state variables, the application of dynamic programming for solving the multi-stage program is not tractable. Let Itt be the discount factor of period t, discounted to period t, i.e.

Itt ¼

t Y

ð1 þ ik ðshortÞÞ;

t > t; Itt ¼ 1;

k¼tþ1

where we use the short rate at time t, it(short), for discounting. Let L1 be the set of funding instruments available at time T1. Funding obtained at time T1 may mature or be called during the first partition (i.e. before or at time T2), during the second partition (i.e. after time T2 and before or at time T3), or during the third partition (i.e. after time T3 and before or at time T4). We denote the set of funding instruments issued at time T1 and matured or called during the first partition of the planning horizon o as L112 , the set of funding instruments issued at time T1 and matured or called during the o o second partition of the planning horizon as L122 3 , and the set of funding instruments issued at time T1 and matured or called during the third partition of the planning horizon o o o o o oo o o as L132 3 4 . Clearly, L1 ¼ L112 [ L122 3 [ L132 3 4 , for each ðo2 ; o3 ; o4 Þ 2 fS2  S3  S4 g. Similarly, we denote the set of funding instruments issued at time T2 and matured or o o called during the second partition of the planning horizon as L222 3 , and the set of funding instruments issued at time T2 and matured or called during the third partition of the o o o oo o o o planning horizon as L232 3 4 . Clearly, L2 ¼ L222 3 [ L232 3 4 , for each ðo2 ; o3 ; o4 Þ 2 fS2  S3  S4 g. Finally, we denote the set of funding instruments issued at time T3 and matured or called during the third partition of the planning horizon as L33. Clearly, o o o L3 ¼ L33 ¼ L332 3 4 , for each ðo2 ; o3 ; o4 Þ 2 fS2  S3  S4 g. o For all funding instruments ‘1 2 L1 2 initiated at time t
0 that mature or are called during the first partition, we obtain the net present values

o

2 ¼ r‘1 ð11Þ

‘ o T2 X Pt 1 2 o

t¼0

I0t 2

;

o o

for all funding instruments ‘1 2 L122 3 initiated at time t
0 that mature or are called during the second partition, we obtain the net present values

o ;o

2 3 ¼ r‘1 ð12Þ

1

‘o T3 X Pt 1 3 o

o

I0T22

t¼T2 þ1

o oo

IT2 t3

;

and all initial funding instruments ‘1 2 L132 3 4 , initiated at time t
0 that mature or are called during the third partition, we obtain the net present values

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

o ;o ;o4

¼

2 3 r‘1 ð13Þ

1

1

o

o

‘ o T4 X Pt 1 4

I0T22 IT2 T3 3

o

IT3 t4

t¼T3 þ1

j

141

:

o o

For all funding instruments ‘2 2 L222 3 , initiated at time t
T2, that mature or are called during the second partition, we obtain the net present values ‘o T3 X Pt 2 3

1

o ;o

2 3 ¼ r‘2 ð22Þ

o

o

I0T22

IT2 t3

t¼T2 þ1

;

o o

and for all funding instruments ‘2 2 L232 3 , initiated at time t
T2, that mature or are called during the third partition, we obtain the net present values o ;o ;o4

2 3 r‘2 ð23Þ

¼

1

1

o

o

‘ o T4 X Pt 2 4

I0T22 IT2 T3 3

o

t¼T3 þ1

IT3 t4

:

We obtain for all initial funding ‘ L33 initiated at time t
T3 the net present values o2 ;o3 ;o4 r‘3 ð33Þ

¼

1

1

o

o

I0T22 IT2 T3 3

o o

o

‘ o T4 X Pt 3 4 o

t¼T3 þ1

o oo

IT3 t4

:

oo

o o o

oo o

2 3 2 3 4 2 3 2 3 4 2 3 4 2 Let N
|S|. Let R‘o1 ¼ r‘1 ð11Þ þ r‘1 ð12Þ þ r‘1 ð13Þ , R‘o2 ¼ r‘2 ð22Þ þ r‘2 ð23Þ , and R‘o3 ¼ r‘3 ð33Þ . Let x‘1 be the amount of funding in instrument ‘1 2 L1 issued at time t
T1, x‘2 be the amount of funding in instrument ‘2 2 L2 issued at time t
T2, and x‘3 be the amount of funding in instrument ‘3  L3 issued at time t
T3. Based on the computation of the net present values, we optimize the funding mix solving the multi-stage (stochastic) linear program:

min s.t.

E v o ¼ v; X x‘1 ¼ 1; ‘1 2L1



X o

x‘1 þ



o o ‘1 2L122 3

X o

x‘1 

R‘o1 x‘1 þ

‘1 2L1

o

x‘2 2 ¼ 0;

‘2 2L2

‘1 2L112

X

X

X

o

x‘2 2 þ

o o ‘2 2L222 3

X

v þ w  u;

o o3

x‘3 2

¼ 0;

‘3 2L3

o

R‘o2 x‘2 2 þ

‘2 2L2 o

X

X

o o3

 w o ¼ 0;

o

o o

R‘o3 x‘3 2

‘3 2L3 o

E w  r;

x‘1 ; x‘2 2 ; x‘3 2 3 ;

v o  0;

P where E w o ¼ ð1=N Þ w o is the estimate of the expected net present value and E v o ¼ P ð1=N Þ v o is the estimate of the risk. As in the single-stage model before, the parameter r is a pre-specified value for the expected net present value of the portfolio. Starting with r ¼ rmax ,

142 j CHAPTER 7

the maximum value of r that can be assumed without the linear program becoming infeasible, we trace out an efficient frontier by successively reducing r from r ¼ rmax to r
0 and computing for each level of r the corresponding value of risk v by solving the multi-stage stochastic linear program. The quantity r , the maximum expected net present value without considering risk, can be obtained by solving the linear program max E w o ¼ r max ; X x‘1 ¼ 1; s.t. ‘1 2L1



X o

x‘1 þ

X o o ‘1 2L122 3

X ‘1 2L1

o

x‘2 2 ¼ 0;

‘2 2L2

‘1 2L112



X

x‘1 

X o o ‘2 2L222 3

X

R‘o1 x‘1 þ

X

o

x‘2 2 þ

o o3

x‘3 2

¼ 0;

‘3 2L3 o

R‘o2 x‘2 2 þ

‘2 2L2

X

o o3

R‘o3 x‘3 2

 w o ¼ 0;

o

o o3

x‘1 ; x‘2 2 ; x‘3 2

 0:

‘3 2L3

Note that the model formulation presented above does not consider the calling of callable bonds as subject to optimization at the decision stages; rather the calling of callable bonds is handled through the calling rule as part of the simulation. Optimizing also the calling of callable bonds at the decision stages requires only a minor extension to the model formulation, but this is not discussed here. A variant of the multi-stage model arises by trading off expected net present value and risk in the objective with l as the risk-aversion coefficient: min lE v o  E w o ; X x‘1 ¼ 1; s.t. ‘1 2L1



X o

x‘1 þ



o o3

R‘o1 x‘1 þ

‘1 2L1 o

o o3

o

x‘2 2 þ

X

o

R‘o2 x‘2 2 þ

‘2 2L2 o

X

o o3

x‘3 2

¼ 0;

‘3 2L3

‘2 2L222

v þ w  u;

7.5

X

x‘1 

‘1 2L122

X

o

x‘2 2 ¼ 0;

‘2 2L2

‘1 2L112

X

X

X

o o3

R‘o3 x‘3 2

 w o ¼ 0;

‘3 2L3

o o o x‘1 ; x‘2 2 ; x‘3 2 3 ; v o

 0:

DURATION AND CONVEXITY

Since the payments from a mortgage pool are not constant, indeed the prepayments depend on the interest rate term structure and its history since the inception of the pool,

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

143

an important issue arises as to how the net present value (price) of the mortgage pool changes as a result of a small change in interest rates. The same issue arises for all funding instruments, namely, how bond prices change as a result of small changes in yield. This is especially interesting in the case of callable bonds. In order to calculate the changes in expected net present value due to changes in interest rates, one usually resorts to first- and second-order approximations, where the first-order (linear, or delta) approximation is called the duration and the second-order (quadratic, or gamma) approximation is called the convexity. While the duration and convexity of non-callable bonds could be calculated analytically, the duration and convexity of a mortgage pool and callable bonds can only be estimated through simulation. We use the terms effective duration and effective convexity to refer to magnitudes estimated through simulation. 7.5.1

Effective Duration and Convexity

Let p¼

T X Ptpool t¼0

It

be the net present value of the payments from the mortgage pool, where Ptpool ¼ at þ bt and It is the discount factor using the short rate for discounting. We compute pv, for scenarios v S, using Monte Carlo simulation, and we calculate the expected net present value (price) of the payments of the mortgage pool as p ¼

1 X o p : N o2S

Note that the payments Ptpool ¼ Ptpool ðik ; k ¼ 0; . . . ; tÞ depend on the interest rate term structure and its history up to period t, where it denotes the vector of interest rates for different maturities at time t. Writing explicitly the dependency, p ¼ pðit ; t ¼ 0; . . . ; T Þ: We now define pþ ¼ pðit þ D; t ¼ 0; . . . ; T Þ as the net present value of the payments of the mortgage pool for an upward shift of all interest rates by D), and p ¼ pðit  D; t ¼ 0; . . . ; T Þ as the net present value of the payments of the mortgage pool for a downward shift of all interest rates by D), where D is a shift of, say, one percentage point in the entire term structure at all periods t
1,. . .,T.

144 j CHAPTER 7

Using the three points p , p, pþ , and the corresponding interest rate shifts D, 0, D, we compute the effective duration of the mortgage pool, as dur ¼

p  pþ ; 2Dp

and the effective convexity of the mortgage pool, con ¼

p þ pþ  2p : 100D2 p

The quantities of the effective duration and effective convexity represent a local first-order (duration) and second-order (duration and convexity) Taylor approximation of the net present value of the payments of the mortgage pool as a function of interest. The approximation considers the effects on a constant shift of the entire yield curve across all points t, t
1,. . ., T. The number 100 in the denominator of the convexity simply scales the resulting numbers. The way it is computed, we expect a positive value for the duration, meaning that decreasing interest rates result in a larger expected net present value and increasing interest rates result in a smaller expected net present value. We also expect a negative value for the convexity, meaning that the function of price versus yield is locally concave. In an analogous fashion we compute the duration durl and the convexity conl for all funding instruments ‘. Let p‘ ¼

m‘ X dt‘ t¼0

It

be the net present value of the payments of the bond ‘, where dt‘ represents the payments of bond ‘ until maturity or until it is called. We compute p‘o using Monte Carlo simulation over v  S, and we calculate the expected net present value (price) of the payments for the bond ‘ as p‘ ¼

1 X N

p‘o :

o2S

We calculate p‘þ ¼ p‘ ði t þ D; t ¼ 0; . . . ; T Þ; and p‘ ¼ p‘ ði t  D; t ¼ 0; . . . ; T Þ; the expected net present values for an upwards and downwards shift of interest rates, respectively. Analogously to the mortgage pool, we obtain the duration of bond ‘ as

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

dur‘ ¼

j

145

p‘  p‘þ ; 2Dp‘

and its convexity as con‘ ¼

p‘ þ p‘þ  2p‘ : 100D2 p‘

Since in the case of non-callable bonds the payments dt‘ are fixed, changes in expected net present value due to changes in interest rates are influenced by the discount factor only. For non-callable bonds we expect a positive value for duration, meaning that increasing interest rates imply a smaller bond value and decreasing interest rates imply a larger bond value. We expect a positive value for convexity, meaning that the function of expected net present value versus interest rates is locally convex. In the case of callable bonds the behavior of the function of expected net present value versus interest rates is influenced not only by the discount rate but also by the calling rule. If interest rates decrease, the bond may be called and the principal returned. The behavior of callable bonds is similar to that of mortgage pools in that we expect a positive value for duration and a negative value for convexity. 7.5.2

Traditional Finance: Matching Duration and Convexity

Applying methods of traditional finance, one would hedge interest rate risk by constructing a portfolio with a duration and a convexity of close to zero, respectively, thus achieving that the portfolio would exhibit no change in expected net present value (price) due to a small shift in the entire yield curve. Duration and convexity matching is also referred to as immunization (see, for example, Luenberger (1998) or as delta and gamma hedging). In the situation of funding mortgage pools, hedging is carried out in such a way that a change in the price of the mortgage pool is closely matched by the negative change in the price of the funding portfolio, such that the change of the total portfolio (mortgage pool and funding) is close to zero. Thus, the duration of the total portfolio is close to zero. In addition, the convexity of the mortgage pool is matched by the (negative) convexity of the funding portfolio, such that the convexity of the total portfolio (mortgage pool and funding) is close to zero. We write the corresponding duration and convexity hedging model as max X

X

r‘ x‘

‘

dur‘ x‘  dg ¼ dur;

‘

X

con‘ x‘  cg ¼ con;

‘

X ‘

x‘ ¼ 1;

x‘  0;

146 j CHAPTER 7

and dgmax  dg  dgmax ;

cgmax  cg  cgmax ;

and expected net present value is maximized in the objective. The variable dg accounts for the duration gap, cg accounts for the convexity gap, dgmax represents a predefined upper bound on the absolute value of the duration gap, and cgmax a predefined upper bound on the absolute value of the convexity gap. Usually, when a duration and convexity hedged strategy is implemented, the model needs to be revised over time as the mortgage pool and the yield curve changes. In practice, updating the funding portfolio may be done on a daily or monthly basis to reflect changes in the mortgage pool due to prepayments and interest rate variations. The duration and convexity hedging model is a deterministic model, uncertainty is considered as a shift of the entire yield curve, and hedged to the extent of the effect of the remaining duration and convexity gap. 7.5.3

Duration and Convexity in the Single-Stage Model

In the single-stage case we add the duration and convexity constraints X

dur‘ x‘  dg ¼ dur;

‘

X

con‘ x‘  cg ¼ con;

‘

and dg max  dg  dg max ;

cg max  cg  cg

to the single-stage linear program using the formulation in which expected net present value and risk are traded off in the objective. Setting the risk aversion l to zero, only the expected net present value is considered in the objective, and the resulting single-stage stochastic program is identical to the duration and convexity hedging formulation from traditional finance as discussed in the previous section. This formulation allows one to constrain the absolute value of the duration and convexity gap to any specified level, to the extent that the single-stage stochastic program remains feasible. By varying the duration and convexity gap, we may study the effect of the resulting funding strategy on expected net present value and risk. 7.5.4

Duration and Convexity in the Multi-Stage Model

In the multi-stage model we wish to constrain the duration and convexity gap not only in the first stage, but at any decision stage and in any scenario. Thus, in the four-stage model discussed above we have one pair of constraint for the first stage, NS2N pairs of constraints in the second stage, and NS2S3N pairs of constraints in the third stage. Accordingly, we

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

147

need to compute the duration and the convexity for the mortgage pool and all funding instruments at any decision point in all stages from one to three. In order to simplify the presentation, we omit the scenario indices v2  S2 and ðo2 ; o3 Þ 2 fS2  S3 g. Let durt and cont be the duration and convexity of the mortgage pool in each stage t, and dur‘t and con‘t be the duration and convexity of the funding instruments issued at each period t. Let durFt and conFt be symbols used to conveniently present the duration and convexity P of the funding portfolio at each period t. In the first stage, durF1 ¼ ‘1 2L1 dur‘1 x‘1 . In the P P second stage, durF2 ¼ ‘2 2L2 dur‘2 x‘2 þ ‘1 2L12 dur‘1 ð12Þ x‘1 , where dur‘1 ð12Þ represents the duration of the funding instruments from the first stage that are still available in the P P P second stage. In the third stage, durF3 ¼ ‘3 2L3 dur‘3 x‘3 þ ‘2 2L23 dur‘2 ð23Þ x‘2 þ ‘1 2L13 dur‘1 ð13Þ x‘1 , where dur‘2 ð23Þ represents the duration of the funding instruments issued in the second stage still available in the third stage, and dur‘1 ð13Þ represents the duration of the funding instruments from the first stage still available in the third stage. P Analogously, in the first stage, conF1 ¼ ‘1 2L1 con‘1 x‘1 . In the second stage, P P conF2 ¼ ‘2 2L2 con‘2 x‘2 þ ‘1 2L12 con‘1 ð12Þ x‘1 , where con‘1 ð12Þ is the convexity of the funding instruments from the first stage still available in the second stage. In the third stage, P P P conF3 ¼ ‘3 2L3 con‘3 x‘3 þ ‘2 2L23 con‘2 ð23Þ x‘2 þ ‘1 2L13 con‘1 ð13Þ x‘1 , where con‘2 ð23Þ is the convexity of the funding instruments issued in the second stage still available in the third stage, and con‘1 ð13Þ is the convexity of the funding instruments from the first stage still available in the third stage. Thus, in the multi-stage case, we add in each stage t ¼ 1; . . . ; 3 and in each scenario v2  S2 and ðo2 ; o3 Þ 2 fS2  S3 g the constraints durFt  dgt ¼ durt ; conFt  cgt ¼ cont ; and dgtmax  dgt  dgtmax ;

cgtmax  cgt  cgtmax :

The variables dgt , cgt account for the duration and convexity gap, respectively, in each decisions stage and in each of the scenarios v2  S2 and ðo2 ; o3 Þ 2 fS2  S3 g. At each decision node the absolute values of the duration and convexity gap are constrained by dgt and cgt , respectively. This formulation allows one to constrain the duration and convexity gap to any specified level, even at different levels at each stage, to the extent that the multistage stochastic linear program remains feasible. Using the formulation of the multi-stage model, in which expected net present value and risk are traded off in the objective, and setting the risk aversion coefficient l to zero, we obtain a model in which the duration and convexity are constrained at each node of the scenario tree and expected net present value is maximized. Since the scenario tree represents simulations of possible events of the future, the model results in a duration and convexity hedged funding strategy, where duration and convexity are constrained at the

148 j CHAPTER 7

decision points of the scenario tree, but not constrained at points in between, where the funding portfolios are carried forward using the decision rules. We use this as an approximation for other duration and convexity hedged strategies, in which duration and convexity are matched, for example, at every month during the planning horizon. One could, in addition, apply funding rules in the simulation that result in duration and convexity hedged portfolios at every month, but such rules have not been applied in the computations used for this chapter. We are now in the position to compare the results from the multi-stage stochastic model trading off expected net present value and downside risk with the deterministic duration and convexity hedged model on the same scenario tree. The comparison looks at expected net present value and risk (represented by various measures), as well as underlying funding strategies.

7.6

COMPUTATIONAL RESULTS

7.6.1

Data Assumptions

For the experiment we used three data sets, based on different initial yield curves, labeled ‘Normal,’ ‘Flat’ and ‘Steep.’ The data represent assumptions about the initial yield curve, the parameters of the interest rate model, and the prepayment function, assumptions about the funding instruments, assumptions about refinancing and the calling rule, and the planning horizon and its partitioning. Table 7.1 presents the initial yield curve (corresponding to a zero coupon bond) for each data set. For each data set the mortgage contract rate is assumed to be one percentage point above the 10-year rate (labeled ‘y10’). For the experiment we consider 16 different funding instruments. Table 7.2 presents the maturity, the time after which the instrument may be called, and the initial spread over the corresponding zero coupon bond (of the same maturity) for each instrument and for each of the data sets. For example, ‘y03nc1’ refers to a callable bond with a maturity of 3 years (36 months) and callable after 1 year (12 months). Corresponding to the data set ‘Normal,’ it could be issued initially (at time t
0) at a rate of TABLE 7.1 Initial Yield Curves Interest rate ()) Label

Maturity (months)

m03 m06 y01 y02 y03 y05 y07 y10 y30

3 6 12 24 36 60 84 120 360

Normal

Flat

Steep

5.18 5.31 5.55 5.97 6.12 6.33 6.43 6.59 6.87

5.88 6.38 6.76 7.06 7.36 7.56 7.59 7.64 7.76

2.97 3.16 3.36 4.18 4.58 5.56 5.97 6.36 7.20

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

149

TABLE 7.2 Spreads for Different Funding Instruments Spread ()) Label

Maturity (months)

m03n m06n y01n y02n y03n y03nc1 y05n y05nc1 y05nc3 y07n y07nc1 y07nc3 y10n y10nc1 y10nc3 y30n

3 6 12 24 36 36 60 60 60 84 84 84 120 120 120 360

Callable after (months)

12 12 36 12 36 12 36

Normal

Flat

Steep

0.17 0.13 0.02 0.00 0.04 0.36 0.08 0.65 0.32 0.17 0.90 0.60 0.22 1.10 0.87 0.30

0.15 0.10 0.08 0.11 0.18 0.61 0.21 0.84 0.41 0.22 0.95 0.73 0.29 1.28 0.94 0.32

0.19 0.13 0.08 0.1 0.12 0.22 0.13 0.22 0.18 0.15 0.45 0.35 0.22 0.57 0.48 0.27

6:12) þ 0:36) ¼ 6:48), where 6.12) is the interest rate from Table 7.1 and 0.36) is the spread from Table 7.2. We computed the results for a pool of $100M. As the target for risk, u, we used the maximum expected net present value obtained using single-stage optimization, i.e. we considered risk as the expected net present value below this target, defined for each data set. In particular, the target for risk equals u
10.5M for the ‘Normal’ data set, u
11.3M for the ‘Flat’ data set and u
18.6M for the ‘Steep’ data set. We first used single-stage optimization using N
300 interest rate paths. These results are not presented here. Then, in order to more accurately compare single-stage and multistage optimization, we used a tree with N
4000 paths, where the sample sizes in each stage are |S2|
10, |S3|
20 and |S4|
20. For this tree the multi-stage linear program has 8213 rows, 11 377 columns and 218 512 non-zero elements. The program can easily be solved on a modern personal computer in a very small amount of (elapsed) time. Also, the simulation runs to obtain the coefficients for the linear program can easily be carried out on a modern personal computer. 7.6.2

Results for the Single-Stage Model

As a base case for the experiment we computed the efficient frontier for each of the data sets using the single-stage model. Figure 7.2 presents the result for the ‘Normal’ data set in comparison with the efficient frontiers obtained from the multi-stage model. (The singlestage results obtained from optimizing on the tree closely resemble the efficient frontiers obtained from optimizing on 300 interest rate paths.) The efficient frontiers for the ‘Flat’ and ‘Steep’ data sets are similar in shape to that of the ‘Normal’ data set. While ‘Normal’ and ‘Flat’ have the typical shape one would expect, i.e. steep at low levels of risk and

150 j CHAPTER 7 Efficient Frontier, Normal 14.0

Exp. Return ($M NPV)

12.0 10.0 8.0 6.0 4.0 2.0 0.0 0.0

1.0

4.0

2.0 3.0 Risk ($M NPV) Single-stage

5.0

Multi-stage

FIGURE 7.2 Efficient frontier for the Normal data set, single- versus multi-stage. TABLE 7.3 Funding Strategies from the Single-Stage Model, Case 95) of Maximum Expected Return, for All Three Data Sets Model

Normal

Flat

Stage 1

Allocation j1m06n

j1y07n

0.853

0.147

Stage 2 Scenario 1 2 3 4 5 6 7 8 9 10

Steep

j1m03n 0.911 Allocation

j2m03n

j2m03n

0.853 0.853 0.853 0.853 0.853 0.853 0.853 0.853 0.853 0.853

0.911 0.911 0.911 0.911 0.911 0.911 0.911 0.911 0.911 0.911

j1y07n

j1m03n

0.089

1.000 j2m03n 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000

bending more flat with increasing risk, it is interesting to note that the efficient frontier for data set ‘Steep’ is very steep at all levels of risk. As a base case we look at the optimal funding strategy at the point of 95) of the maximum expected net present value. In the graphs of the efficient frontiers this is the second point down from the point of the maximum expected net present value point. The funding strategies for all three data sets are presented in Table 7.3. The label ‘j1’ in front of the funding acronyms means that the instrument is issued in stage 1, and the label ‘j2’ refers to the instrument’s issuance in stage 2. For example, a funding instrument called

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

151

‘j1y07n’ is a non-callable bond with a maturity of 7 years issued at stage 1. While the single-stage model naturally does not have a second stage, variables with the prefix ‘j2’ for the second stage represent the amount of each instrument to be held at stage 2 according to the decision rules. This is to facilitate easy comparisons between the single-stage and multi-stage funding strategies. In the ‘Normal’ case, the optimal initial funding mix consists of 85.3) six-month noncallable debt and 14.7) seven-year non-callable debt. The risk level of this strategy is 3.1M NPV. In the ‘Flat’ case, the optimal initial funding mix consists of 91.1) threemonth non-callable debt and 8.9) seven-year non-callable debt. The risk level of this strategy is 3.1M NPV. In the ‘Steep’ case, the initial funding mix consists of 100.0) threemonth non-callable debt (short-term debt). The risk level of this strategy is 2.6M NPV. 7.6.3

Results for the Multi-Stage Model

Figure 7.2 presents the efficient frontier for the ‘Normal’ data set. In addition to the multistage efficient frontier, the graph also contains the corresponding single-stage efficient frontier for better comparison. The results for the ‘Flat’ and ‘Steep’ data sets are very similar in shape to that of the ‘Normal’ data set and therefore are not presented graphically. The results show substantial differences in the risk and expected net present value profile of multi-stage versus single-stage funding strategies. For any of the three data sets, the efficient frontier obtained from the multi-stage model is significantly north-west of that from the single-stage model, i.e. multi-stage optimization yields a larger expected net present value at the same or smaller level of risk. In all three data cases, we cannot compare the expected net present values from the single-stage and multi-stage model at the same level of risk, because the risk at the minimum risk point of the single-stage model is larger than the risk at the maximum risk point of the multi-stage model. In the ‘Normal’ case, the minimum risk of the single-stage curve is about 2.7M NPV. Since the efficient frontier is very steep at low levels of risk, we use the point of 85) of the maximum risk as the ‘lowest risk’ point, even if the risk could be further decreased by an insignificant amount. The maximum expected net present value point of the multi-stage curve has a risk of about 2.4M NPV. At this level of risk, the expected net present value on the single-stage efficient frontier is about 8.9M NPV, versus the expected net present value on the multi-stage curve of about 13.6M NPV, which represents an improvement of 52.8). In the ‘Flat’ case, we compare the point with the smallest risk of 3.0M NPV on the single-stage efficient frontier with that with the largest risk of 2.5M NPV on the multi-stage efficient frontier. The expected net present value at the two points is 13.6M NPV (multi-stage) versus 9.6M NPV (single-stage), which represents an improvement of 41.7). In the ‘Steep’ case, we compare the point with the smallest risk of 2.6M NPV on the single-stage efficient frontier with that with the largest risk of 1.9M NPV on the multi-stage efficient frontier. The expected net present value at the two points is 20.4M NPV (multi-stage) versus 18.6M NPV (single-stage), which represents an improvement of 9.7). As for the single-stage model, we look at the funding strategies at the point of 95) of the maximum expected return (on the efficient frontier the second point down from the

152 j CHAPTER 7 TABLE 7.4 Funding Strategy from the Multi-Stage Model, Case 95) of Maximum Expected Return, Normal Data Set Stage 1

Allocation j1m06n

j1y05nc1

0.785

0.215

Stage 2 Scenario 1 2 3 4 5 6 7 8 9 10

Allocation j2m03n

j2y03n

j2y03nc1

j2y05nc1

j2y10n

j2y10nc1

j2y10nc3

0.597

0.188

j2y30n 1.000

0.125

0.420

0.240

0.508

0.423

0.785 0.069

0.785 0.785 0.251 0.301 0.785

0.035

0.749 0.261

0.188

maximum expected return point). The funding strategies are presented in Table 7.4 for the ‘Normal’ data set, and in Table 7.A1 and Table 7.A2 of Appendix 7.A for the ‘Steep’ and ‘Flat’ data sets, respectively. In the ‘Normal’ case, the initial funding mix consists of 78.5) six-month non-callable debt, and 21.5) five-year debt callable after one year. After one year the 78.5) six-month debt (that according to the decision rules is refinanced through short-term debt and is for disposition in the second stage) and, if called in certain scenarios, also the five-year callable debt are refunded through various mixes of short-term debt: three-year, five-year and ten-year callable and non-callable debt. In one scenario, labeled ‘1,’ in which interest rates fall to a very low level, the multi-stage model resorts to funding with 30-year noncallable debt in order to secure the very low rate for the future. The risk associated with this strategy is 1.85M NPV and the expected net present value is 12.9M NPV. The corresponding (95) of maximum net present value) strategy of the single-stage model, discussed above, has a risk of 3.1M NPV and a net present value of 10.0M NPV. Thus, the multi-stage strategy exhibits 57.4) of the risk of the single-stage strategy and a 29) larger expected net present value. In the ‘Flat’ case, the initial funding mix consists of 50.9) three-month non-callable debt and 49.1) six-month non-callable debt. After one year the entire portfolio is refunded through various mixes of short-term, three-year, and ten-year callable and noncallable debt. The multi-stage strategy exhibits 68) of the risk of the single-stage strategy and a 20.6) larger expected net present value. In the ‘Steep’ case, the initial funding consists of 100) three-month non-callable debt. After one year the portfolio is refunded through various mixes of short-term, three-year and five-year non-callable, and ten-year callable and non-callable debt. The multi-stage strategy exhibits 62) of the risk of the single-stage strategy, and a 4.3) larger expected net present value.

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

153

Summarizing, the results demonstrate that multi-stage stochastic programming potentially yields significantly larger net present values at the same or even lower levels of risk, with significantly different funding strategies, compared with single-stage optimization. Using multi-stage stochastic programming for determining the funding of mortgage pools promises to lead, in the average, to significant profits compared with using single-stage funding strategies. 7.6.4

Results for Duration and Convexity

Funding a mortgage pool by a portfolio of bonds that matches the (negative) value of duration and convexity, the expected net present value of the total of mortgage pool and bonds is invariant to small changes in interest rates. However, duration and convexity give only a local approximation, and the portfolio needs to be updated as time goes on and interest rates change. The duration and convexity hedge is one-dimensional, since it considers only changes of the whole yield curve by the same amount and does not consider different shifts for different maturities. The multi-stage stochastic programming model takes into account multi-dimensional changes of interest rates and considers (via sampling) the entire distribution of possible yield curve developments. In this section we quantify the difference between duration and convexity hedging versus hedging using the single- and multi-stage stochastic programming models. Table 7.5 gives the initial (first stage) values for duration and convexity (as obtained from the simulation runs) for the mortgage pool and for all funding instruments for each of the three yield curve cases ‘Normal,’ ‘Flat’ and ‘Steep.’ In each of the three yield curve cases the mortgage pool exhibits a positive value for duration and a negative value for convexity. All funding instruments have positive values for duration, non-callable bonds exhibit a positive value TABLE 7.5 Initial Duration and Convexity Normal

Flat

Steep

Label

Dur.

Conv.

Dur.

Conv.

Dur.

Conv.

mortg. m03n m06n y01n y02n y03n y03nc1 y05n y05nc1 y05nc3 y07n y07nc1 y07nc3 y10n y10nc1 y10nc3 y30n

3.442 0.248 0.492 0.971 1.882 2.737 1.596 4.278 1.914 3.274 5.622 2.378 3.406 7.335 1.555 3.409 13.711

1.880 0.001 0.003 0.010 0.038 0.081 0.384 0.205 0.810 0.012 0.367 0.128 0.269 0.656 0.246 0.160 2.880

2.665 0.247 0.491 0.964 1.861 2.686 1.376 4.160 1.613 3.068 5.448 1.671 3.117 7.083 1.397 3.265 13.306

1.548 0.001 0.003 0.010 0.038 0.079 0.186 0.197 0.700 0.111 0.351 0.959 0.298 0.624 0.052 0.346 2.743

2.121 0.249 0.495 0.982 1.917 2.801 1.338 4.378 1.478 3.114 5.769 0.898 3.122 7.538 0.593 3.086 14.204

2.601 0.001 0.003 0.011 0.039 0.084 0.235 0.212 0.632 0.241 0.380 0.032 0.230 0.681 0.119 0.446 3.011

154 j CHAPTER 7

for convexity, and callable bonds show a negative value for convexity. Note as an exception the positive value for convexity of bond ‘y07nc1.’ To both the single-stage and multi-stage stochastic programming models we added constraints that, at any decision point, the duration and convexity of the mortgage pool and the funding portfolio are as close as possible. Maximizing expected return by setting the risk aversion l to zero, we successively reduced the gap in duration and convexity between the mortgage pool and the funding portfolio. We started from the unconstrained case (the maximum expected return maximum risk case from the efficient frontiers discussed above) and reduced first the duration gap and subsequently the convexity gap, where we understand as duration gap the absolute value of the difference in duration between the funding portfolio and the mortgage pool, and as convexity gap the absolute value of the difference in convexity between the funding portfolio and the mortgage pool. We will discuss the results with respect to the downside risk measure (expected value of returns below a certain target), as discussed in Section 7.2.9, and also with respect to the standard deviation of the returns. We do not discuss the results for duration and convexity obtained from the single-stage model and focus on the more interesting multi-stage case. 7.6.5

Duration and Convexity, Multi-Stage Model

Figures 7.3 and 7.4 give the risk return profile for the ‘Normal’ case with respect to downside risk and standard deviation, respectively. We compare the efficient frontier (already depicted in Figure 7.2) obtained from minimizing downside risk for different levels of expected return (labeled ‘Downside’) with the risk return profile obtained from restricting the duration and convexity gap (labeled ‘Delta–Gamma’). For different levels of duration and convexity gap, we maximized expected return. The unconstrained case with respect to duration and convexity is identical to the point on the efficient frontier with the maximum expected return. Figure 7.3 shows that the downside risk eventually increases Risk-Return Profile, Multi, Normal

16

Exp. Return ($M NPV)

14 12 10 8 6 4 2 0 0

1

2 3 Risk ($M NPV) Downside

4

Delta-Gamma

FIGURE 7.3 Risk-return profile, multi-stage model, Normal data set, risk as downside risk.

5

j

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

155

Risk-Return Profile, Multi, Normal 16

Exp. Return ($M NPV)

14 12 10 8 6 4 2 0 0

1

2

3

5 6 4 Risk STD ($M NPV) Downside

7

8

9

10

Delta-Gamma

FIGURE 7.4 Risk-return profile, multi-stage model, Normal data set, risk as standard deviation.

when decreasing the duration and convexity gap and was significantly larger than the minimized downside risk on the efficient frontier. We are especially interested in comparing the point with the smallest duration and convexity gap with the point of minimum downside risk on the efficient frontier. The point with the smallest downside risk on the efficient frontier exhibits expected return of 9.5M NPV and a downside risk of 1.5M NPV. The point with the smallest duration and convexity gap has expected return of 7.9M NPV and a downside risk of 3.1M NPV. The latter point is characterized by a maximum duration gap in the first and second stage of 0.5 and of 1.0 in the third stage, and by a maximum convexity gap of 2.0 in the first and second stage and of 4.0 in the third stage. A further decrease of the convexity gap was not possible as it led to an infeasible problem. The actual duration gap in the first stage was 0.5, where the negative value indicates that the duration of the funding portfolio was smaller than that of the mortgage pool, and the actual convexity gap in the first stage was 1.57, where the positive value indicates that the convexity of the funding portfolio was larger than that of the mortgage pool. Comparing the points with regard to their performance, restricting the duration and convexity gap led to a decrease in expected return of 15) and an increase of downside risk by a factor of 2. It is interesting to note that, for the point on the efficient frontier with the smallest risk, the first-stage duration gap was 0.97 and the first-stage convexity gap was 1.75. Looking at the risk in terms of standard deviation of returns, both minimizing downside risk and controlling the duration and convexity gap led to smaller values of standard deviation. In the unconstrained case, the smallest standard deviation was 3.3M NPV obtained at the minimum downside risk point, and the smallest standard deviation in the constrained case (3.8M NPV) was obtained when the duration and convexity gap was smallest. Table 7.6 gives the funding strategy for the minimum downside risk portfolio and Table 7.7 the funding strategy for the duration and convexity constrained case at

156 j CHAPTER 7 TABLE 7.6 Funding Strategy, Case Minimum Downside Risk, Multi-Stage Model, Normal Data Set Stage 1

Allocation j1m06n

j1y05n

j1y05nc1

0.266

0.235

0.218

j2m03n

j2y01n

j2y03nc1

Stage 2 Scenario 1 2 3 4 5 6 7 8 9 10

0.017

0.266 0.031 0.266 0.118 0.012 0.266

j1y05nc3 0.281 Allocation j2y05n 0.148

j2y10n

j2y10nc1

j2y10nc3

0.102

0.165

0.240

j2y30n 0.079

0.117

0.149

0.032

0.237

0.185

0.014 0.300 0.201

0.252 0.019 0.053

0.047

minimum gap. Both strategies use six-month non-callable debt, five-year non-callable debt and five-year debt callable after one year for the initial funding. The minimum downside risk portfolio used in addition five-year debt callable after three years, while the duration and convexity constrained case used seven-year non-callable debt and ten-year non-callable debt. In the second stage, funding differed significantly across the different scenarios for both funding strategies. In the duration and convexity constrained case, the multi-stage optimization model resorted to calling five-year debt callable after one year in scenario ‘4’ at a fraction and in scenario ‘10’ at the whole amount. In the minimum risk case, calling of debt other than by applying the decision rules did not happen. In the second stage, interest rates were low in scenarios ‘1’ and ‘8’ and were high in scenarios ‘4,’ ‘6’ and ‘10.’ The minimum downside risk strategy tended towards more long-term debt when interest rates were low and towards more short-term debt when interest rates were high. The amounts for each of the scenarios depended on the dynamics of the process and the interest rate distributions. The duration and convexity constrained strategy was, of course, not in the position to take advantage of the level of interest rates, and funding was balanced to match the duration and convexity of the mortgage pool. The results for the case of the ‘Flat’ and ‘Steep’ yield curves are very similar. For the ‘Flat’ case, the maximum duration gap in the first and second stage was set to 0.5 and to 1.0 in the third stage, and the maximum convexity gap was set to 2.0 in the first and second stage and to 4.0 in the third stage. The duration and convexity gap could not be decreased further since the problem became infeasible. The actual duration gap in the first stage was 0.5, and the actual convexity gap in the first stage was 1.5. In Appendix 7.A, Table 7.A3 gives the initial funding and the second-stage updates for the minimum downside risk portfolio, and Table 7.A4 gives the funding strategy for the duration and convexity constrained case. For the ‘Steep’ case, the maximum duration gap in the first and second stage was set to 0.5 and was set to 1.0 in the third stage, and the maximum

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

157

TABLE 7.7 Funding Strategy, Case Duration and Convexity Constrained, Multi-Stage Model, Normal Data Set Stage 1

Allocation j1m06n

j1y05n

j1y05nc1

j1y07n

j1y10n

0.122

0.083

0.534

0.239 Allocation

0.022

j2y03nc1

j2y05n

Stage 2 Scenario 1 2 3 4 5 6 7 8 9 10 Stage 2 ]Scenario 1 2 3 4 5 6 7 8 9 10

j2m03n

j2m06n

j2y01n

j2y05nc1

0.466

j2y07nc1 0.117

0.064 0.506

0.043

0.044

0.064 0.567 0.072 0.316

0.340

Allocation

Call

]j2y10n

]j2y10nc1

0.011 0.058 0.152 0.058 0.059 0.122 0.089

0.112

]j2y10nc3

j1y05nc1

0.073

0.029 0.004

0.051 0.534

convexity gap was set to 3.0 in the first and second stage and to 6.0 in the third stage. Further decrease made the problem become infeasible. The actual duration gap in the first stage was 0.5, and the actual convexity gap in the first stage was 1.6. In Appendix 7.A, Table 7.A5 gives the initial funding and the second-stage updates for the minimum downside risk portfolio, and Table 7.A6 gives the funding strategy for the duration and convexity constrained case. Again, in both the ‘Flat’ and ‘Steep’ case, one could see in the minimum downside risk case a tendency to use longer-term debt when interest rates were low and shorter-term debt when interest rates were high, and in the duration and convexity constrained case, funding was balanced to match the duration and convexity of the mortgage pool. Summarizing, in each case reducing the duration and convexity gap helped control the standard deviation of net present value but did nothing to reduce downside risk. Multistage stochastic programming led to larger than or equal expected net present value at each level of risk (both downside risk and standard deviation of net present value). There may be reasons for controlling the duration and convexity gap in addition to controlling

158 j CHAPTER 7 Duration Gap versus Interest Rate 10.00 8.00

Duration Gap (%)

6.00 4.00 2.00 0.00

–2.00 –4.00 –6.00 0.00

2.00

4.00

95% max expo. ret.

6.00 Interest Rate (%) Dur. and conv constr.

8.00

10.00

12.00

Min. downside risk

FIGURE 7.5 Duration gap versus interest rate, multi-stage, Normal data set.

downside risk via the multi-stage stochastic programming model. For example, a conduit might like the results of the stochastic programming model, but not wish to take on too much exposure regarding duration and convexity gap. In the following we will therefore compare runs of the multi-stage model with and without duration and convexity constraints. Figure 7.5 sheds more light on the cause of the performance gain of the multi-stage model versus the duration and convexity hedged strategy. The figure presents the duration gap (the difference of the duration of the funding portfolio and the mortgage pool) versus the interest rate (calculated as the average of the yield curve) for each of the second-stage scenarios of the data set ‘Normal.’ When interest rates are very low, the 95) maximum expected return strategy takes on a significant positive duration gap to lock in the low rates for a long time. It takes on a negative duration gap when interest rates are high, in order to remain flexible, should interest rates fall in the future. The minimum downside risk strategy exhibits a similar pattern, but less extreme. Thus, the multi-stage model makes a bet on the direction interest rates are likely to move, based on the information about the interest rate process. In contrast, the duration and convexity constrained strategy cannot take on any duration gap (represented by the absolute value of 0.5 at which the gap was constrained) and therefore must forsake any gain from betting on the likely direction of interest rates. 7.6.6

Out-of-Sample Simulations

In order to evaluate the performance of the different strategies in an unbiased way, true out-of-sample evaluation runs need to be performed. Any solution at any node in the tree obtained by optimization must be evaluated using an independently sampled set of observations.

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

159

For the single-stage model this evaluation is rather straightforward. Having obtained an optimal solution from the single-stage model (using sample data set one), we run simulations again with a different seed. Using the new independent sample (data set two), we start the optimizer again, however now with the optimal solution fixed at the values as obtained from the optimization based on sample data set one. Using the second set of observations of data set two we calculate risk and expected returns. To independently evaluate the results obtained from a K-stage model (having n
K1 sub-periods), we need K independent sets of K-stage trees of observations. We describe the procedure for K
4. Using data set one we solve the multi-stage optimization problem and obtain an optimal first-stage solution. We simulate again (using a different seed) over all stages to obtain data set two. Fixing the optimal first-stage solution at the value obtained from the optimization based on data set one, we optimize based on data set two and obtain a set of optimal second-stage solutions. We simulate again (with a different seed) to obtain independent realizations for stages three and four, thereby keeping the observations for stage two the same as in data set two, and obtain data set three. Fixing the first-stage decision at the level obtained from the optimization using data set one and all second-stage decisions at the level obtained from the optimization based on data set two, we optimize again to obtain a set of optimal third-stage decisions. We simulate again (using a different seed) to obtain independent outcomes for stage four, thereby keeping the observations for stage two and three the same as in data sets two and three, respectively, and obtain data set four. Fixing the first-stage decision, all second-stage decisions, and all third-stage decisions at the levels obtained from the optimization based on data sets one, two and three, respectively, we finally calculate risk and returns based on data set four. The out-of-sample evaluations resemble how the model could be used in practice. Solving the multi-stage model (based on data set one), an optimal first-stage solution (initial portfolio) would be obtained and implemented. Then one would follow the strategy (applying the decision rules) for 12 months until decision stage two arrives. At this point, one would re-optimize, given that the initial portfolio had been implemented and that particular interest rates and prepayments had occurred (according to data set two). The optimal solution for the second stage would be implemented, and one would follow the strategy for four years until decision stage three arrives. At this point, one would re-optimize, given that the initial portfolio and a second stage update had been implemented and that particular interest rates and prepayments had occurred (according to data set three). The optimal solution for the third stage would be implemented, and one would follow the strategy (applying the decision rules) until the end of the horizon (according to data set four). Now one possible path of using the model has been evaluated. Decisions had no information about particular outcomes of future interest rates and prepayments, and were computed based on model runs using data independent from the observed realization of the evaluation simulation. Alternatively, one could simulate a strategy involving re-optimization every month, but this would take significantly more computational effort with likely only little to be gained.

160 j CHAPTER 7 Out-of-Sample Efficient Frontier, Normal

Expected Return ($M NPV)

14.00 12.00 10.00 8.00 6.00 4.00 2.00 0.00 0.00

1.00 Single-stage

2.00 3.00 Risk ($M NPV) Multi-stage

4.00

5.00

Multi-stage dur. constr.

FIGURE 7.6 Out-of-sample risk /return profile, multi-stage model, Normal data set, risk as downside risk.

Using the out-of-sample evaluation procedure, we obtain N
|S| out-of-sample simulations of using the model as discussed in the above paragraph, and we are now in the position to derive statistics about out-of-sample expected returns and risk. Figure 7.6 presents the out-of-sample efficient frontiers for the ‘Normal’ data set and downside risk. The figure gives the out-of-sample efficient frontier for the multi-stage model without duration and convexity constraints, the out-of-sample efficient frontier when the duration gap was constrained to be less than or equal to 1.5), and the out-ofsample efficient frontier for the single-stage model. The out-of-sample evaluations demonstrate clearly that the multi-stage model gives significantly better results than the single-stage model. For example, the point with the maximum expected returns of the multi-stage model gave expected returns of 10.2M NPV and a downside risk of 3.5M NPV. The minimum risk point on the single-stage out-of-sample efficient frontier gave expected returns of 7.6M NPV, and a downside risk of also 3.5M NPV. Thus, at the same level of downside risk the multi-stage model gave 34) higher expected returns. The minimum downside risk point of the multi-stage model gave expected returns of 9.0M NPV and a downside risk of 2.1M NPV. Comparing the minimum downside risk point of the multistage model with that of the single-stage model, the multi-stage model had 19.2) larger expected returns at 61) of the downside risk of the single-stage model. The efficient frontier of the duration-constrained strategy was slightly below that without duration and convexity constraints. Figure 7.7 gives the out-of-sample riskreturn profiles when measuring risk in terms of standard deviation. In this setting the multi-stage strategy performed significantly better than the single-stage strategy at every level of risk, where the difference was between 17.6) and 23.6). The duration-constrained strategy did not span as wide a range in risk as the unconstrained strategy. But for the risk points attained by the constrained strategy, the unconstrained strategy achieved a somewhat higher expected return.

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

161

Out-of-Sample Risk-Return Profile, Normal, Larger Sample 14.00

Expected Return ($M NPV)

12.00 10.00 8.00 6.00 4.00 2.00 0.00 0.00

1.00

2.00

3.00

Single-stage

4.00 5.00 6.00 Risk STD ($M NPV) Multi-stage

7.00

8.00

9.00

10.00

Multi-stage dur. constr.

FIGURE 7.7 Out-of-sample risk /return profile, multi-stage model, Normal data set, risk as standard deviation.

The out-of-sample evaluations for the ‘Flat’ and ‘Steep’ data sets gave very similar results qualitatively. The results are included in Appendix 7.A. For the ‘Flat’ data set, Figure 7.A1 presents out-of-sample efficient frontiers for downside risk and Figure 7.A2 the risk return profile for risk in standard deviations. For the ‘Steep’ data set, Figure 7.A3 presents out-of-sample efficient frontiers for downside risk and Figure 7.A4 the risk return profile for risk in standard deviations for the data set ‘Steep.’ For both data sets, the multi-stage model gave significantly better results; the downside risk was smaller and expected returns were larger. 7.6.7

Larger Sample Size

All results discussed so far were obtained from solving a model with a relatively small number of scenarios at each stage, i.e. |S2|
10, |S3|
20 and |S4|
20, with a total of 4000 scenarios at the end of the forth stage. This served well for analysing and understanding the behavior of the multi-stage model in comparison with the single-stage model and with Gamma and Delta hedging. Choosing a larger sample size will improve the obtained strategies (initial portfolio and future revisions) and therefore result in better (out-of-sample simulation) results. Of course, the accuracy of prediction of the models will be improved also. In order to show the effect of using a larger sample size, we solved and evaluated the models using a sample size of 24 000 scenarios, i.e. |S2|
40, |S3|
30 and |S4|
20. The results for the data set ‘Normal’ are presented in Figure 7.8 for downside risk and in Figure 7.9 for risk as standard deviation. Indeed, one can see improved performance in both smaller risk and larger expected NPV compared with the smaller sample size (compare with Figures 7.6 and 7.7). The point with the maximum expected returns for the multi-stage model gave expected returns of 12.9M NPV and a downside risk of 2.4M NPV. The minimum risk point for the

162 j CHAPTER 7 Out-of-Sample Efficient Frontier, Normal, Larger Sample 14.00

Expected Return ($M NPV)

12.00 10.00 8.00 6.00 4.00 2.00 0.00 0.00

1.00

2.00

4.00

3.00

5.00

Risk ($M NPV) Single-stage

Multi-stage

Multi-stage dur. constr.

FIGURE 7.8 Larger sample, out-of-sample risk /return profile for the multi-stage model, Normal data set, risk as downside risk.

Out-of-Sample Risk-Return Profile, Normal, Larger Sample 14.00

Expected Return ($M NPV)

12.00 10.00 8.00 6.00 4.00 2.00 0.00 0.00

1.00

2.00

3.00

4.00

5.00

6.00

7.00

8.00

9.00

10.00

Risk STD ($M NPV) Single-stage

Multi-stage

Multi-stage dur. constr.

FIGURE 7.9 Larger sample, out-of-sample risk /return profile for the multi-stage model, Normal data set, risk as standard deviation.

single-stage model had expected returns of 9.0M NPV and a downside risk of 2.7M NPV. So, at slightly smaller downside risk the multi-stage model gave expected returns that were 43) higher. The minimum downside risk point of the multi-stage model had expected returns of 10.9M NPV and a downside risk of 1.5M NPV. Comparing the minimum downside risk points of the multi-stage and single-stage models, the multi-stage model has 18.4) larger expected returns at 57) of the downside risk of the single-stage model. Again, the efficient frontier of the duration-constrained strategy was slightly below that

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

163

Predicted vs. Out-of-Sample Efficient Frontier, Normal, Larger Sample 14.00

Expected Return ($M NPV)

12.00 10.00 8.00 6.00 4.00 2.00 0.00 0.00

1.00

2.00 3.00 Risk ($M NPV)

Multi-stage out-of-sample

4.00

5.00

Multi-stage predicted

FIGURE 7.10 Larger sample, predicted versus out-of-sample efficient frontier for the multi-stage model, Normal data set.

without duration and convexity constraints. Measuring risk in terms of standard deviation, results are similar to those when using the small sample size: the multi-stage strategy performed significantly better than the single-stage strategy at every level of risk, with a difference in expected returns between 23.3) and 24.0). Again, the durationconstrained strategy exhibited smaller risk (in standard deviations) at the price of slightly smaller expected returns. Figure 7.10 gives a comparison of the multi-stage efficient frontiers predicted (insample) versus evaluated out-of-sample. One can see that when using the larger sample size of 24,000, the predicted and the out-of-sample evaluated efficient frontier look almost identical, thus validating the model. It is evident that using larger sample sizes will result in both better performance and a more accurate prediction.

7.7

LARGE-SCALE RESULTS

For the actual practical application of the proposed model, we need to consider a large number of scenarios in order to obtain small estimation errors regarding expected returns and risk and accordingly stable results. We now explore the feasibility of solving large-scale models. Table 7.8 gives measures of size for models with larger numbers of scenarios. For example, model L4 has 80,000 scenarios at the fourth stage, composed of |S2|
40, |S|
40 and |S4|
50, and model L5 has 100,000 scenarios at the fourth stage composed of |S2|
50, |S3|
40 and |S4|
50; both models have a sufficiently large sample size in each stage. In the case of problem L5 with 100,000 scenarios the corresponding linear program had 270,154 constraints, 249,277 variables, and 6,291,014 non-zero coefficients. Table 7.9 gives

164 j CHAPTER 7 TABLE 7.8 Large-Scale Models, Dimensions Scenarios Model L1 L2 L3 L4 L5 L20

Problem Size

Stage 2

Stage 3

Stage 4

Total

Rows

Columns

Non-zeros

10 20 30 40 50 200

40 40 40 40 40 40

50 50 50 50 50 50

20,000 40,000 60,000 80,000 100,000 400,000

54,034 108,064 162,094 216,124 270,154 808,201

49,918 99,783 149,623 199,497 249,277 931,271

1,249,215 2,496,172 3,756,498 5,021,486 6,291,014 14,259,216

the elapsed times for simulation and optimization, obtained on a Silicon Graphics Origin 2000 workstation. While the Origin 2000 at our disposition is a multi-processor machine with 32 processors, we did not use the parallel feature, and all results were obtained using single processor computations. For the direct solution of the linear programs, we used CPLEX (1989–1997) as the linear programming optimizer. We contrast the results to using DECIS (Infanger 1989–1999), a system for solving stochastic programs, developed by the author. Both CPLEX and DECIS were accessed through GAMS (Brooke et al. 1988). DECIS exploited the special structure of the problems and used dual decomposition for their solution. The problems were decomposed into two stages, breaking between the first and the second stage. The simulation runs for model L5 took less than an hour. The elapsed solution time solving the problem directly was 13 hours and 28 minutes. Solving the problem using DECIS took significantly less time, 2 hours and 33 minutes. Encouraged by the quick solution time using DECIS, we generated problem L20 with 400,000 scenarios (composed of |S2|
200, |S3|
40 and |S4|
50) and solved it in 11 hours and 34 minutes using DECIS. Problem L20 had 808,201 constraints, 931,271 variables, and 14,259,216 non-zero coefficients. Using parallel processing, the simulation times and the solution times could be reduced significantly. Based on our experiences with parallel DECIS, using six processors we would expect the solution time for the model L5 with 100,000 scenarios to be less than 40 minutes, and using 16 processors one could solve model L20 with 400,000 scenarios in about one hour. We actually solved a version of the L3 model with 60,000 scenarios, composed of |S2|
50, |S3|
40 and |S4|
30, in less than 5 minutes using parallel DECIS on 16 processors. TABLE 7.9 Large-Scale Models, Solution Times Model L1 L2 L3 L4 L5 L20

Scenarios

Simul. time (s)

Direct sol. time (s)

20,000 40,000 60,000 80,000 100,000 400,000

688 1390 2060 2740 3420

1389.75 6144.56 14 860.54 28 920.69 48 460.02

Decomp. sol. time (s) 1548.91 5337.76 9167.47 41 652.28

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

7.8

j

165

SUMMARY

The problem of funding mortgage pools represents an important problem in finance faced by conduits in the secondary mortgage market. The problem concerns how to best fund a pool of similar mortgages through issuing a portfolio of bonds of various maturities, callable and non-callable, and how to refinance the debt over time as interest rates change, prepayments are made and bonds mature. This chapter presents the application of stochastic programming in combination with Monte Carlo simulation for effectively and efficiently solving the problem. Monte Carlo simulation was used to estimate multiple realizations of the net present value of all payments when a pool of mortgages is funded initially with a single bond, where pre-defined decision rules were applied for making decisions not subject to optimization. The simulations were carried out in monthly time steps over a 30-year horizon. Based on a scenario tree derived from the simulation results, a single-stage stochastic programming model was formulated as a benchmark. A multi-stage stochastic programming model was formulated by splitting up the planning horizon into multiple sub-periods, representing the funding decisions (the portfolio weights and any calling of previously issued callable bonds) for each sub-period, and applying the pre-defined decision rules between decision points. In order to compare the results of the multi-stage stochastic programming model with hedging methods typically used in finance, the effective duration and convexity of the mortgage pool and of each funding instrument was estimated at each decision node, and constraints bounding the duration and convexity gap to close to zero were added (at each node) to the multi-stage model to approximate a duration and convexity hedged strategy. An efficient method for obtaining an out-of-sample evaluation of an optimal strategy obtained from solving a K-stage stochastic programming model was presented, using K independent (sub-)trees for the evaluation. For different data assumptions, the efficient frontier of expected net present value versus (downside) risk obtained from the multi-stage model was compared with that from the single-stage model. Under all data assumptions, the multi-stage model resulted in significantly better funding strategies, dominating the single-stage model at every level of risk, both in-sample and by evaluating the obtained strategies via out-of-sample simulations. Also, for all data assumptions, the out-of-sample simulations demonstrated that the multi-stage stochastic programming model dominated the duration and convexity hedged strategies at every level of risk. Constraining the duration and convexity gap reduced risk in terms of the standard deviation of net present value at the cost of a smaller net present value, but failed in reducing the downside risk. The results demonstrate clearly that using multi-stage stochastic programming results in significantly larger profits, both compared with using single-stage optimization models and with using duration and convexity hedged strategies. The multi-stage model is better than the single-stage model because it has the option to revise the funding portfolio according to changes in interest rates and pre-payments, therefore reflecting a more realistic representation of the decision problem. It is better than the duration and convexity hedged strategies because it considers the entire distribution of the yield curve

166 j CHAPTER 7

represented by the stochastic process of interest rates, compared with the much simpler hedge against a small shift of the entire yield curve as used in the duration and convexity hedged strategies. Small models with 4000 scenarios and larger ones with 24,000 scenarios were used for determining the funding strategies and the out-of-sample evaluations. The out-of-sample efficient frontiers of the larger models were shown to be very similar to the (in-sample) predictions, indicating a small optimization bias. The larger models were solved in a very short (elapsed) time of a few minutes. Large-scale models with up to 100,000 scenarios were shown to solve in a reasonable elapsed time using decomposition on a serial computer, and in a few minutes on a parallel computer.

ACKNOWLEDGEMENTS Funding for research related to this chapter was provided by Freddie Mac. The author is grateful to Jan Luytjes for many valuable discussions on the subject. Editorial comments by three anonymous referees and by John Stone on previous versions of this paper helped enhance the presentation.

REFERENCES Bellman, R., Dynamic Programming, 1957 (Princeton University Press: Princeton, NJ). Brooke, A., Kendrik, D. and Meeraus, A., GAMS, A Users Guide, 1988 (Scientific Press: South San Francisco, CA). Carino, D.R., Kent, T., Myers, D.H., Stacy, C., Sylvanus, M., Turner, A.L., Watanabe, K. and Ziemba, W.Y., The Russel Yasuda Kasai model: An asset-liability model for a Japanese insurance company using multistage stochastic programming. Interfaces, 1994, 24, 29  49. Cox, J.C., Ingersoll, J.E. and Ross, S.A., A theory of the term structure of interest rates. Econometrica, 1985, 53, 385407. CPLEX Optimization, Inc., Using the CPLEX Callable Library, 1989 1997 (CPLEX Optimization, Inc.: Incline Village, NV). Dantzig, G.B. and Glynn, P.W., Parallel processors for planning under uncertainty. Ann. Oper. Res., 1990, 22, 1 21. Dembo, R. and Mausser, H., The put/call efficient frontier. Algo. Res. Q., 2000, 3, 13 25. Dempster, M.A.H. and Thorlacius, A.F., Stochastic simulation of international economic variables and asset returns, the falcon asset model in 8th International AFIR Colloquium, 1998, pp. 2945. Di Domenica, N., Valente, P., Lucas, C.A. and Mitra, G., A review of scenario generation for stochastic programming and simulation, CARISMA Research Report CTR 31, to appear in The State of the Art in Planning under Uncertainty (Stochastic Programming), edited by G.B. Dantzig and G. Infanger, 2006 (Kluwer: Dordrecht). Frauendorfer, K. and Schu¨ rle, M., Refinancing mortgages in Switzerland. Applications of Stochastic Programming (MPS-SIAM Series on Optimization), edited by S.W. Wallace and W.T. Ziemba, 2005 (SIAM: Philadelphia). Golub, B., Holmer, M., McKendal, R., Pohlman, L. and Zenios, S.A., A stochastic programming model for money management. Eur. J. Oper. Res., 1995, 85, 282 296. Hiller, R.S. and Eckstein, J., Stochastic dedication: Designing fixed income portfolios using massively parallel benders decomposition. Mgmt Sci., 1993, 39, 14221438.

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

167

Ho, T.S.Y. and Lee, S.-B., Term structure movements and pricing interest rate contingent claims. J. Finan., 1986, XLI, 10111029. Hull, J. and White, A., Pricing interest rate securities. Rev. Finan. Stud., 1990, 3, 573592. Infanger, G., Monte Carlo (importance) sampling within a benders decomposition algorithm for stochastic linear programs. Ann. Oper. Res., 1992, 39, 65 95. Infanger, G., Planning Under Uncertainty—Solving Large-Scale Stochastic Linear Programs, 1994 (Boyd and Fraser) (The Scientific Press Series: Danvers, MA). Infanger, G., Managing risk using stochastic programming, In INFORMS National Meeting, 1996. Infanger, G., DECIS User’s Guide, 1989 1999 (Infanger Investment Technology: Mountain View, CA). Infanger, G., Dynamic asset allocation strategies using a stochastic dynamic programming approach. In Handbook of Asset and Liability Management, edited by S. Zenios and W.T. Ziemba, 2006 (Elsevier: Amsterdam). Infanger, G. and Morton, D., Cut sharing for multistage stochastic linear programs with interstage dependency. Math. Program., 1996, 75, 241 256. Kang, P. and Zenios, S., Complete prepayment models for mortgage-backed securities. Mgmt Sci., 1992, 38, 16651685. Kusy, M.I. and Ziemba, W.T., A bank asset and liability management model. Oper. Res., 1986, 35, 356 376. Lekkas, V. and Luytjes, J., Estimating a prepayment function for 15- and 30-year fixed rate mortgages. FHLMC Technical Report, 1993; Rates model enhancement summary. Freddie Mac Internal Document. Longstaff, F.A. and Schwarz, E.S., Interest rate volatility and the term structure: A two-factor general equilibrium model. J. Finan., 1992, XLVII, 12591282. Luenberger, D.G., Investment Science, 1998 (Oxford University Press: New York). Luytjes, J., Estimating a yield curve/inflation process. FHLMC Technical Report, 1993. Luytjes, J., Memorandum FR call rule—2Q96 release. FHLMC Technical Report, 1996. Markowitz, H., Portfolio selection. J. Finan., 1952, 7, 77 91. Mulvey, J.M. and Vladimirou, H., Stochastic network programming for financial planning problems. Mgmt Sci., 1992, 38, 1642 1664. Nielsen, S. and Zenios, S.A., A stochastic programming model for funding single-premium deferred annuities. Math. Program., 1996, 75, 177 200. Prindiville, M., Advances in the application of stochastic programming to mortgage finance. PhD thesis, Department of Engineering—Economic Systems and Operations Research, Stanford University, CA, 1996. Vasicek, O.A., An equilibrium characterization of the term structure. J. Finan. Econ., 1977, 5, 177 188. Wallace, S.W. and Ziemba, W.T., Applications of Stochastic Programming, (MPS-SIAM Series on Optimization), 2005 (SIAM: Philadelphia). Zenios, S.A., A model for portfolio management with mortgage-backed securities. Ann. Oper. Res., 1993, 43, 337356. Ziemba, W.T., The Stochastic Programming Approach to Asset, Liability, and Wealth Management, 2003 (The Research Foundation of AIMR: Charlottesville, VA).

168 j CHAPTER 7

APPENDIX 7.A: TABLES AND GRAPHS FOR DATA SETS ‘FLAT’ AND ‘STEEP’

TABLE 7.A1 Funding Strategy from the Multi-Stage Model, Case 95) of Maximum Expected Return, Flat Data Set Stage 1

Allocation j1m03n

j1m06n

0.509

0.491

Stage 2 Scenario 1 2 3 4 5 6 7 8 9 10

Allocation j2m03n

j2y03n

j2y03nc1

j2y10n

j2y10nc1

j2y10nc3

0.397

j2y30n 0.603

1.000 0.251 1.000

0.590

0.011 0.514

1.000 0.984 0.620 0.521 1.000

0.248

0.148 0.238 0.016

0.380 0.151

0.328

TABLE 7.A2 Funding Strategy from the Multi-Stage Model, Case 95) of Maximum Expected Return, Steep Data Set Stage 1

Allocation j1m03n 1.000

Stage 2 Scenario 1 2 3 4 5 6 7 8 9 10

Allocation j2m03n

j2y03n

j2y03nc1

0.038 0.717

j2y05n

j2y10n

0.296

0.544

j2y10nc1

0.122

0.283 1.000

1.000 0.745 1.000 1.000 0.609 0.973 1.000

j2y30n

0.085

0.170

0.272 0.027

0.119

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

169

TABLE 7.A3 Funding Strategy from the Multi-Stage Model, Case Minimum Downside Risk, Flat Data Set Stage 1

Allocation j1m06n

j1y05nc3

0.698

0.302

Stage 2 Scenario 1 2 3 4 5 6 7 8 9 10

Allocation j2m03n

j2m06n

1 2 3 4 5 6 7 8 9 10

j2y03nc1

0.046 0.201 0.698 0.119 0.698 0.603 0.333 0.302 0.698

0.292 0.145

Stage 2 Scenario

j2y03n

Allocation j2y10n

j2y10nc1

j2y10nc3

0.129

j2y30n 0.053

0.039

0.698 0.039

0.221

0.214

0.280 0.095

0.095 0.085 0.301

0.127

j2y05n

j2y05nc3

0.418

0.052

170 j CHAPTER 7 TABLE 7.A4 Funding Strategy from the Multi-Stage Model, Case Duration and Convexity Constrained, Flat Data Set Stage 1

Allocation j1m06n

j1y05n

j1y05nc1

j1y07n

0.448

0.225

0.201

0.125 Allocation

Stage 2 Scenario 1 2 3 4 5 6 7 8 9 10

j2m03n

j2m06n

j2y02n

0.567 0.102 0.370

j2y03n

Call j2y03nc1

j2y10nc1

j1y05nc1

0.082 0.346 0.078 0.251

0.033

0.164

0.018

0.631 0.397

0.634 0.448 0.277

j2y10n

0.051 0.036 0.016

0.412

0.372

0.201

TABLE 7.A5 Funding Strategy from the Multi-Stage Model, Case Minimum Downside Risk, Steep Data Set Stage 1

Allocation j1m03n 1.000

Stage 2 Scenario 1 2 3 4 5 6 7 8 9 10

Allocation j2m03n

j2y01n

j2y03n

j2y03nc1

0.299 0.644

j2y05n

j2y10n

0.389

0.215

j2y10nc1

j2y30n 0.097

0.356 1.000

1.000 0.734 1.000 1.000 0.600 0.958 1.000

0.005

0.088

0.178

0.246 0.042

0.149 8.23454E-4

5.126743E-4

j

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

171

TABLE 7.A6 Funding Strategy from the Multi-Stage Model, Case Duration and Convexity Constrained, Steep Data Set Stage 1

Allocation j1y01n

j1y03n

j1y10n

0.682

0.305

0.013

Stage 2

Allocation

Scenario

j2m03n

1 2 3 4 5 6 7 8 9 10

j2y01n

j2y02n

j2y03n

j2y03nc1

j2y10n

0.598 0.049

j2y30n 0.084

0.053 0.633

0.629 0.531

0.151 0.682

0.158 0.500

0.524 0.182

0.682 0.682 0.392

0.289

Out-of-Sample Efficient Frontier, Flat 14.00

Expected Return ($M NPV)

12.00 10.00 8.00 6.00 4.00 2.00 0.00 0.00

1.00 Single-stage

2.00 3.00 Risk ($M NPV) Multi-stage

4.00

5.00

Multi-stage dur. constr.

FIGURE 7.A1 Out-of-sample risk /return profile for the multi-stage model, Flat data set, risk as downside risk.

172 j CHAPTER 7 Out-of-Sample Risk-Return Profile, Flat 14.00

Expected Return ($M NPV)

12.00 10.00 8.00 6.00 4.00 2.00 0.00 0.00

1.00

2.00

3.00

Single-stage

4.00 5.00 6.00 Risk STD ($M NPV) Multi-stage

7.00

8.00

9.00

10.00

Multi-stage dur. constr.

FIGURE 7.A2 Out-of-sample risk /return profile for the multi-stage model, Flat data set, risk as standard deviation.

Out-of-Sample Efficient Frontier, Steep 20.00

Expected Return ($M NPV)

18.00 16.00 14.00 12.00 10.00 8.00 6.00 4.00 2.00 0.00 0.00

1.00 Single-stage

2.00 3.00 Risk STD ($M NPV) Multi-stage

4.00

5.00

Multi-stage dur. constr.

FIGURE 7.A3 Out-of-sample risk /return profile for the multi-stage model, Steep data set, risk as downside risk.

STOCHASTIC PROGRAMMING FOR FUNDING MORTGAGE POOLS

j

173

Out-of-Sample Risk-Return Profile, Steep 20.00

Expected Return ($M NPV)

18.00 16.00 14.00 12.00 10.00 8.00 6.00 4.00 2.00 0.00 0.00

1.00

2.00

3.00

Single-stage

4.00 5.00 6.00 Risk STD ($M NPV) Multi-stage

7.00

8.00

9.00

10.00

Multi-stage dur. constr.

FIGURE 7.A4 Out-of-sample risk /return profile for the multi-stage model, Steep data set, risk as standard deviation.

CHAPTER

8

Scenario-Generation Methods for an Optimal Public Debt Strategy MASSIMO BERNASCHI, MAYA BRIANI, MARCO PAPI and DAVIDE VERGNI CONTENTS 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 8.2 Problem Description and Basic Guidelines . . . . . . . . . . . . . . . . . . . . . . 176 8.3 Models for the Generation of Interest Rate Scenarios . . . . . . . . . . . . . . . 177 8.3.1 Simulation of the ECB Official Rate . . . . . . . . . . . . . . . . . . . . . . . 179 8.3.2 Simulation of the Fluctuations . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 8.3.3 Principal Component Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 181 8.3.4 Multivariate Mean-Reverting Models . . . . . . . . . . . . . . . . . . . . . . 183 8.4 Validation of the Simulated Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . 189 8.5 Conclusions and Future Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 Appendix 8.A: Estimation with Discrete Observations . . . . . . . . . . . . . . . . . . 194

8.1

T

INTRODUCTION

is of paramount importance for any country. Mathematically speaking, this is a stochastic optimal control problem with a number of constraints imposed by national and supranational regulations and by market practices. The Public Debt Management Division of the Italian Ministry of Economy decided to establish a partnership with the Institute for Applied Computing in order to examine this problem from a quantitative viewpoint. The goal is to determine the composition of the portfolio issued every month that minimizes a predefined objective function (Adamo et al. 2004), which can be described as an optimal combination between cost and risk of the public debt service. Since the main stochastic component of the problem is represented by the evolution of interest rates, a key point is to determine how various issuance strategies perform under HE MANAGEMENT OF THE PUBLIC DEBT

175

176 j CHAPTER 8

different scenarios of interest rate evolution. In other words, an optimal strategy for the management of the public debt requires a suitable modeling of the stochastic nature of the term structure of interest rates. Note that hereafter we do not report how to forecast the actual future term structure, but how to generate a set of realistic possible scenarios. For our purposes, the scenarios should cover a wide range of possible outcomes of future term structures in order to provide a reliable estimate of the possible distribution of future debt charges. This distribution is useful in a risk-management setting to estimate, for instance, a sort of Cost at Risk (CaR) of the selected issuance policy (Risbjerg and Holmund 2005). The chapter is organized as follows. Section 8.2 describes the problem. Section 8.3 describes the models employed for generating future interest rate scenarios. Section 8.4 introduces criteria to validate a scenario. Section 8.5 concludes with future perspectives of this work.

8.2

PROBLEM DESCRIPTION AND BASIC GUIDELINES

It is widely accepted that stock prices, exchange rates and most other interesting observables in finance and economics cannot be forecast with certainty. Interest rates are an even more complex issue because it is necessary to consider the term structure, which is a multi-value observable. Despite this difficulty, there are a number of studies that, from both the academic and practitioner viewpoint, deal with interest rate modeling (for a comprehensive survey of interest rate modeling, see James and Webber (2000)). The most common application of existing term structure models is the evaluation of interest-rate-contingent claims. However, our purpose is different, since we aim to find an optimal strategy for the issuance of Italian public debt securities. In a very simplified form the problem can be described as follows. The Italian Treasury Department issues a number of different types of securities. Securities differ in the maturity (or expiration date) and in the rules for the payment of interest. Short-term securities (those having maturity up to two years) do not have coupons. Medium- and long-term securities (up to 30 years) pay cash dividends, every 6 months, by means of coupons. The problem is to find a strategy for the selection of public debt securities that minimizes the expenditure for interest payment (according to the ESA95 criteria (Jackson 2000)) and satisfies, at the same time, the constraints on debt management. The cost of future interest payments depends on the future value of the term structure (roughly speaking, when a security expires or a coupon is paid, there is the need to issue a new security whose cost depends on the term structure at expiration time). That is the reason why we need to generate scenarios of future interest rates. Adamo et al. (2004) show that, for a set of term structure evolutions and Primary Budget Surplus (PBS) realizations, such an optimization problem can be formulated as a linear programming problem with linear constraints. Broadly speaking, this is a typical risk-management problem: once we find an optimal strategy for a specific realization of the term structure evolution, we need to determine the expenditure for interest if a different scenario takes place. As a consequence, we need to simulate the term structure under the objective measure dynamics. This requirement entails an implicit evaluation of market-price-of-risk dynamics.

j

SCENARIO-GENERATION METHODS

177

6.5 6 5.5 5 January 1999 June 1999 January 2000 June 2000 January 2001 June 2001 January 2002 June 2002 January 2003 June 2003 January 2004 June 2004 January 2005 June 2005

4.5 4 3.5 3 2.5 2

0

50

100

150 200 Maturity (in months)

250

300

350

FIGURE 8.1 Evolution of the term structure in the time frame January 1999
/September 2005.

The first issue to consider in the model selection process is the time frame. For the purposes of the Ministry, a reasonable planning period is 3
5 years. Within such a long period, the term structure may change significantly, as shown in Figure 8.1. Figure 8.2 reports the monthly evolution of the swap interest rates for the following maturities: 3, 6, 12, 24, 36, 60, 120, 180, 360 months, along with the value of the European Central Bank (ECB) official rate and a simple interpolation of such rate in the same period (January 1999
September 2005). The interpolation is obtained by joining two successive jumps of the ECB official rate by means of a line. Such an interpolation mimics the evolution of interest rates, especially for short maturities, and we use it as a simple approximation of the ECB trend. This is our basic data set for the analysis and generation of new scenarios of term structure evolution. We selected swap rates due to their availability for any maturity regardless of the actual issuance of a corresponding security. /

/

8.3

MODELS FOR THE GENERATION OF INTEREST RATE SCENARIOS

From Figure 8.2 it is apparent that any rate (regardless of its maturity) has a strong correlation with the ECB rate. This is not surprising and we use this observation to develop an original approach to the generation of future term-structure scenarios that can be described as a multi-step process: i. generation of a scenario for the future ECB official rate; ii. generation of the fluctuations of each rate with respect to the ECB official rate; iii. validation of the resulting scenario to determine whether it is acceptable or not.

178 j CHAPTER 8 6.5 ECB ECBIN BOT 3 BOT 6 BOT 12 CTZ 24

6 5.5

BTP 36 BTP 60 BTP 120 BTP 180 BTP 360

5 4.5 4 3.5 3 2.5 2 1.5

0

10

20

30 40 50 60 Months (starting on 1/1/1999)

70

80

90

FIGURE 8.2 Evolution of interest rates with maturity from three months up to 30 years in the time frame January 1999
/September 2005. The thick line is a linear interpolation of the ECB official rate (represented by ).

The basic idea is that the ECB official rate represents a ‘reference rate’ and all the other rates are generated by the composition of this ‘reference rate’ plus a characteristic, maturity-dependent, fluctuation. In other words, each rate is determined by the sum of two components: the first component is proportional to the (linearly interpolated) ECB official rate; the second component is the orthogonal fluctuation of that rate with respect to the ECB. In mathematical terms, each rate rh is decomposed as rth ¼ ah rtecb þ pth;? ;

ð8:1Þ

where rtecb is the linear interpolation of the ECB official rate and ah is given by the expression

ah ¼

hr h  r ecb i hr h i  hr ecb i ; hðr ecb Þ2 i hr ecb i2

ð8:2Þ

where
× denotes the average value of the enclosed time series. By construction, the time series pth;? has null correlation with recb. The factors ah are different for each maturity and a larger value of ah indicates a stronger correlation with the ECB official rate. Table 8.1 reports the value of ah for each maturity considered. As expected, longer maturities are less correlated with the ECB official rate.

j

SCENARIO-GENERATION METHODS

179

TABLE 8.1 Value of a for Different Maturities Maturity

ah

Maturity

ah

Maturity

ah

3 months 6 months 12 months

0.9833 0.9514 0.9094

2 years 3 years 5 years

0.7693 0.6695 0.5425

10 years 15 years 30 years

0.3949 0.3545 0.3237

8.3.1

Simulation of the ECB Official Rate

For the simulation of the future ECB official rate we do not employ a macroeconomic model since we assume that the interventions of the ECB can be represented as a stochastic jump process (we return to this point in Section 8.5). Some features of this process are readily apparent by looking at the evolution of the ECB official rate since January 1999 (see Figure 8.2) when the ECB official rate replaced national official rates for the Euro ‘zone’ countries:  



there are, on average, three ECB interventions per year; the ECB rate jumps (with approximately the same probability) by either 25 or 50 basis points; and there is a strong ‘persistence’ in the direction of the jumps. That is, there is a high probability that a rise will be followed by another rise and that a cut will be followed by another cut.

We model the ECB interventions as a combination of two processes: (i) a Poisson process that describes when the ECB changes the official rate; and (ii) a Markov process that describes the sign of the change. Then we resort to an exponential distribution to simulate the waiting time between two changes of the ECB official rate, and to a Markov Chain for simulation of the direction (positive or negative) of the change. The parameter of the exponential distribution can easily be estimated from available data (that is from the waiting times, in months, between the jumps occurring in the past) by means of the Maximum Likelihood Estimation (MLE) method. It turns out to be approximately equal to three months. Since there are two possible states (positive and negative) in the Markov chain, the corresponding transition matrix has four entries (positive
positive, positive
negative, negative
negative, negative
positive). We estimate the values of each entry by looking at the historical data and, in particular, at the probability that a change is in the same, or in the opposite, direction of the previous change. It is interesting to note that the probability of having a change in the same direction of the previous change is pretty high, approximately 85). The width of the jump is selected between two possible values (25 or 50 basis point) with the same probability. In summary, the ECB official rate at time t is defined as /

/

/

/

ECBt ¼ ECB0 þ

Nt X s¼1

as Cs 1;s ;

ð8:3Þ

180 j CHAPTER 8 6 Scenario 1 Scenario 2 Scenario 3 Scenario 4 Scenario 5

Simulated ECB Rate

5

4

3

2

1

0

0

5

10

15 20 25 30 35 Month of the Simulated Time Frame

40

45

50

FIGURE 8.3 Simulations of the future ECB official rate for a planning period of 48 months (the initial value is always the September 2005 official rate, that is 2)).

where Nt is the realization at time t of the Poisson process that represents the total number of jumps up to time t, as is the random width of jump s, and Cs-1,s represents the sign of jump s given the sign of jump s1. Moreover, in order to prevent the ECB official rate from reaching unrealistic values (e.g. negative values), we impose a lower bound on the simulated rate. This lower bound is set equal to a fixed value (currently 1)). Of course, such a value can easily be modified if the evolution of the real ECB rate shows that it is no longer adequate. Any jump of the ECB rate that brings it below the lower bound is discarded. Figure 8.3 shows a few simulations of the future ECB official rate produced using this approach. 8.3.2

Simulation of the Fluctuations

Figure 8.4 shows the result of decomposition (8.1) applied to the data of Figure 8.2 (only the component having null correlation with the ECB rate is shown in the figure). First, we highlight that the historical fluctuations are correlated. This is apparent from the correlation matrix reported in Table 8.2. As a consequence, simulation of their evolution requires multivariate models in order to represent this correlation. In other words, it is not possible to model the dynamics of the fluctuations of a single rate without taking into account the dynamics of the fluctuations of all rates. To this end, we decided to follow two approaches having different features that complement each other. The first approach is based on principal component analysis.

SCENARIO-GENERATION METHODS

j

181

5 ECB ECBN BOT 3 BOT 6

4

BOT 12 CTZ 24 BTP 36 BTP 60

BTP 120 BTP 180 BTP 360

3

2

1

0

–1

0

10

20

30 40 50 Months (starting on January 1999)

60

70

80

FIGURE 8.4 Evolution of the components of interest rates, pth;? , having null correlation with the linearly interpolated ECB rate in the time frame January 1999
/September 2005.

TABLE 8.2 Correlation Matrix of the Components of Each Rate Having Null Correlation with the (Linearized) ECB Official Rate

p3,  p6,  p12, p24, p36, p60, p120, p180, p360,

8.3.3

p3,

p6,

p12,

p24,

p36, 

p60,

p120, 

p180. 

p360,

1.000 0.822 0.625 0.533 0.505 0.490 0.431 0.384 0.311

0.822 1.000 0.937 0.867 0.827 0.779 0.682 0.612 0.520

0.625 0.937 1.000 0.970 0.934 0.877 0.770 0.697 0.610

0.533 0.867 0.970 1.000 0.991 0.956 0.869 0.803 0.717

0.505 0.827 0.934 0.991 1.000 0.986 0.922 0.866 0.786

0.490 0.779 0.877 0.956 0.986 1.000 0.971 0.931 0.862

0.431 0.682 0.770 0.869 0.922 0.971 1.000 0.990 0.949

0.384 0.612 0.697 0.803 0.866 0.931 0.990 1.000 0.981

0.311 0.520 0.610 0.717 0.786 0.862 0.949 0.981 1.000

Principal Component Analysis

Principal component analysis (PCA) is a well-known technique in time series analysis and has been in use for a number of years in the study of fixed income markets (Litterman and Scheinkman 1991). In general, PCA assumes that the underlying process is a diffusion. The data we employ do not have the jump components produced by the ECB interventions thanks to the decomposition procedure (8.1) described previously. From this viewpoint the data appear suitable for PCA. The procedure we follow is the standard one.

182 j CHAPTER 8 



h;? For each de-trended rate pth;? we calculate the differences dht ¼ pth;? pt 1 (t is the time index). We apply PCA to the m time series of the differences, which we indicate by d. This means:
calculating the empirical covariance matrix (Lij ) of d; and
finding a diagonal matrix of eigenvalues Ld and an orthogonal matrix of eigenvectors E  [e1,. . .,em] such that the covariance matrix is factored as Ld ¼ ELhist E T . / /

The eigenvectors with the largest eigenvalues correspond to the components that have the strongest correlation in the data set. For the data set of the monthly yield changes from January 1999 to September 2005, the first three components represent 98.3) of the total variance in the data. The projections of the original data onto the first three principal components do not show autocorrelation phenomena (the covariance matrix is, as expected in the PCA, very close to the Ld matrix). Usually, the first three components are interpreted respectively as (i) a level shift, (ii) a slope change, and (iii) a curvature change. Since the PCA is applied, in our case, to de-trended data (with respect to the ECB official rate) the meaning of the components could be different. However, the plot of the first three components shown in Figure 8.5 does not seem too different from similar studies that consider yield changes directly (Baygun et al. 2000). To create a new scenario for the fluctuations it is necessary to take a linear combination of the principal components. If N is the number of principal components (N3 in the present case), an easy way is to compute d ¼ Fm, where F  [e1,. . . ,eN] is an m N matrix composed of the eigenvectors corresponding to the first N eigenvalues, and m a vector with 0.8 First PC Second PC Third PC

0.6

Loadings

0.4 0.2 0

–0.2 –0.4 –0.6

0

50

100

150 200 Maturity (in months)

250

300

350

FIGURE 8.5 The three most significant components computed from monthly yield changes, January 1999
/ September 2005.

SCENARIO-GENERATION METHODS

j

183

N elements. In principle, the value of each element of this vector could be selected at will. However, there are two common choices: the first is to assign a value taking into account the meaning of the corresponding principal component. For instance, if the purpose is to test the effect of a level shift with no slope or curvature change, it is possible to assign a value only to the first element, leaving the other two elements equal to zero. The other pffiffiffiffiffi ffi choice is to compute d ¼ F Ld Z, where Z is a vector of N independent, normally distributed, variables. Therefore, each element of the vector m is drawn from a normal distribution with variance equal to the eigenvalue corresponding to the principal component. In the present work, we followed the second approach since we did not want to make a priori assumptions. Hence, the evolution of pth;? is described by the following equation:

h;? þ pth;? ¼ pt 1

N X

Fhj

qffiffiffiffi lj Zj ;

ð8:4Þ

j¼1

where lj is the jth eigenvalue. Note that a linear combination of the principal components provides a vector of fluctuations for one time period only. Actually, since the planning period is 3
5 years and the time step is one month, we need, for a single simulation, a minimum of 36 up to a maximum of 60 vectors of fluctuations. Obviously, it is possible to repeat the procedure for the generation of the fluctuations, but, as a result of limited sampling, the covariance matrix of the resulting simulated fluctuations may appear quite different with respect to the covariance matrix of the historical fluctuations. Although a different behavior does not necessarily imply that the simulated fluctuations are meaningless, it is clear that we must keep this difference under control. In Section 8.4 we describe how we dealt with this issue. Figure 8.6 shows the results of a simulation, based on the PCA, of the evolution of interest rates for a planning period of 48 months and a few samples of the corresponding term structures. The results of another simulation reported in Figure 8.7 show how this technique is able to also produce inverted term structures in which long-term or mediumterm maturities have lower returns with respect to short-term maturities. This is not the most common situation, of course, but it may happen, as shown in Figure 8.1 (see the term structure of January 2001). The second approach for the simulation of the fluctuations aims at maintaining a closer relationship with the historical fluctuations and assumes that each fluctuation has its own long-term level to which it tends to return. For this reason, we simulate the fluctuations as a set of mean-reverting processes. /

8.3.4 8.3.4.1

Multivariate Mean-Reverting Models The basic stochastic process A widely used model for the description of interest rate

evolution is based on the following equation proposed by Cox, Ingersoll and Ross (1985) (CIR):

184 j CHAPTER 8 8

7

6

5

4

3

2 0

5

10

15

BCE BTP 36

BCEIN BTP 60

50

100

20

25 Months

BOT 3 BTP 120

30

BOT 6 BTP 180

35

40

BOT 12 BTP 360

45

50

CTZ 24

8.0 7.5 7.0 6.5 6.0 5.5 5.0 4.5 4.0 3.5 0

150 200 Maturity (months)

250

300

6 months

12 months

18 months

24 months

30 months

36 months

42 months

48 months

350

FIGURE 8.6 Simulation of the evolution of interest rates based on the PCA (top panel) and the corresponding term structures (bottom panel).

SCENARIO-GENERATION METHODS

j

185

5

4.5

4

3.5

3

2.5

6 months 12 months 18 months 24 months 30 months 36 months 42 months 48 months

2

1.5

0

50

100

150 200 Maturity (in months)

250

300

350

FIGURE 8.7 Another set of term structures produced by the PCA-based technique. Note how those corresponding to 30, 36, 42 and 48 months are inverted.

pffiffiffiffi drt ¼ kðm rt Þ dt þ s rt dBt ;

ð8:5Þ

where k, m and s are positive constants and rt is the short-term interest rate. The CIR model belongs to the class of mean-reverting and exponential-affine models (Duffie and Kan 1996). This means that the short rate is elastically pulled to the constant long-term value m (mean-reversion) and that the yield to maturity on all bonds can be written as a linear function of the state variables (affine models have been carefully analysed by Duffie et al. 2000). With respect to other, more simple, models belonging to the same class, such as the Vasicek (1977) model, the CIR model guarantees that the interest rates are represented by continuous paths and remain positive if condition 2km > s2 is fulfilled, and r0
0. Moreover, the volatility is proportional to the interest rate. A review of some of the estimation methods for these models is reported in Appendix 8.A. As already mentioned, if we used the simple one-factor model (8.5), we would neglect a fundamental point for the generation of scenarios of the future term structure, that is the correlation among fluctuations corresponding to different maturities. In order to capture this correlation element, in the following we propose a simple multi-dimensional extension of model (5) for the generation of the orthogonal fluctuations with respect to the ECB official rate. We consider the following equation: dpth;?

¼ kh ðmh

pth;? Þ dt

qffiffiffiffiffiffiffiffi X M h;? þ pt shj dB jt ;

for h ¼ 1; . . . ; M

ð8:6Þ

j¼1

on the interval [0, T], T being the time horizon for the generation of scenarios. Here Bt ¼ ðBt1 ; . . . ; BtM Þ is an M-dimensional Brownian motion representing M sources of

186 j CHAPTER 8

randomness in the market, and kh
0, mh
0, and shj are constants such that the matrix s is symmetric and the following conditions hold true: detðsÞ 6¼ 0;

1 kh mh > ðs2 Þhh ; 2

8h ¼ 1; . . . ; M:

ð8:7Þ

It is easy to show that the inequality (8.7) is a sufficient condition for the existence of a unique solution with positive trajectories to the stochastic differential Equation (8.6) M starting at p0 2 ð0; 1Þ ; see, in particular, Duffie and Kan (1996). Since the var
cov pffiffiffiffi > matrix SðpÞSðpÞ , with Si;j ðpÞ¼ pi sij , i; j ¼1; . . . ; M, is not an affine function of p, model (8.6) does not belong to the affine yield class (Duffie et al. 2000) unless s is diagonal, in which case the model reduces to a collection of M uncorrelated CIR univariate processes. However, model (8.6) preserves some features of the class studied in Duffie anf Kan (1996) and Duffie et al. (2000), such as analytical tractability and convenient econometric implementation, while the volatility of each component is proportional only to the component itself. Below, we focus our attention on the problem of the estimation of (8.6). /

8.3.4.2. The estimation of model (8.6) There is a growing literature on estimation methods for term structure models; see, in particular, Ait-Sahalia (2002). A vast literature is specifically devoted to the estimation of affine models (see Balduzzi et al. 1996 and references therein). Here we shall discuss an ad hoc method to estimate model (8.6), which is based on a discrete-time maximum likelihood method (MLE). It is now well recognized that discretization of continuous-time stochastic differential equations introduces a bias in the estimation procedure. However, such a bias is negligible when the data have daily frequency (Bergstrom 1988). For this reason we use daily swap data in this case. In discrete time, the process in Equation (8.6) becomes

pthiþ1 ¼ pthi þ kh ðmh pthi ÞD þ

M qffiffiffiffiffiffiffiffi X j pthi D shj Zi ;

i ¼ 1; . . . ; n 1;

ð8:8Þ

j¼1

for h 1, . . . , M, where ti iD and D Tun. Here, Zi ¼ ðZi1 ; . . . ; ZiM Þ, i1,. . . ,n 1, are independent multivariate normal random variables with zero mean and covariance matrix IM, IM being the identity matrix of order M. We observe that the distributional properties of the process (8.8) depend only on k, m and s2. Therefore, in order to generate scenarios from model (8.8), it suffices to know an estimate of this matrix. For this, we introduce the matrix 1 G ¼ ½s2  . Estimation involves maximizing the log-likelihood function associated with a sequence of observations p^thi > 0: ‘n ¼ z þ

n 1 n 1 1X logðdet GÞ ei> Gei ; 2 2 i¼1

where z ¼ ½ðn 1ÞM logð2pÞ=2 and

ð8:9Þ

SCENARIO-GENERATION METHODS

eih :¼ Ehi þ Zhi kh mh þ chi kh ;

j

187

ð8:10Þ

Ehi , Zhi and chi being constants independent of the parameters of the model, given by p^thiþ1 p^thi ffiffiffiffiffiffiffiffi ; Ehi ¼ q p^thi D

sffiffiffiffiffi D ; Zhi ¼ p^th

chi ¼

qffiffiffiffiffiffiffiffi p^th D; i

ð8:11Þ

i

for every i1,. . . ,n 1 and h 1, . . . , M. We reduce the number of parameters from 2MM2 to 2M  [M(M1)/2] by introducing the following new variables: G ¼ C > C;

ah ¼ kh mh ;

h ¼ 1; . . . ; M;

ð8:12Þ

where C is a lower triangular matrix with strictly positive entries on the main diagonal. (using the Cholesky factorization of G 1 ¼ A> A, with A upper triangular, C ¼ ½A 1 > ). We associate with this matrix a vector c 2 IRMðMþ1Þ=2 according to the relation Ci;j ¼ c½iði 1Þ=2þj ;

8M  i  j  1:

ð8:13Þ

2

2 Using these variables, we have det G ¼ ½det C ¼ PM h¼1 Chh , while (8.9) can be rewritten as follows:

‘n ðc; a; kÞ ¼ z þ ðn 1Þ

M X

logðChh Þ

h¼1

n 1 1X 2 jCei j : 2 i¼1

ð8:14Þ

Therefore, the calibration of model (8.6) can be obtained by computing the maximum of M ln for a; k 2 ð0; 1Þ and c 2 IRMðMþ1Þ=2 , satisfying chðhþ1Þ=2 > 0, for any h 1, . . . , M. We observe that the computation can be reduced to the maximum of U ða; kÞ ¼ sup ‘n ðc; a; kÞ c

M

ð8:15Þ

M

on ð0; 1Þ  ð0; 1Þ . Since ln is concave compared with c, it is easy to show that U is 2M well defined and, for every ða; kÞ 2 ð0; 1Þ , there is a c  ¼ c  ða; kÞ such that U ða; kÞ ¼ ‘n ðc  ; a; kÞ

ð8:16Þ

holds true. The optimizer c  is related, via relation (8.13), to the lower triangular matrix pffiffiffiffiffi ðRij = Rii Þij defined by

188 j CHAPTER 8

0 0 1 Ri1 B BR  C B B i2 C B C ¼ V 1 B i B B .. C @ @ . A

1 0 0 C C .. C C; . A n 1

Rii

i ¼ 1; . . . M;

ð8:17Þ

where >

ðVi Þh1 h2 ¼ ðVi ða; kÞÞh1 h2 :¼ ½e h1  e h2 ;

h1 ; h2 ¼ 1; . . . ; i:

ð8:18Þ

For details of this result, see Papi (2004). Summarizing, it is possible to select an optimal value of C C(a ,k) and then calculate the maximum of the function U which depends only on 2M variables, thus reducing the computational burden. Since we do not know whether the function U is concave or not, it is not possible to resort to gradient methods in order to find a global maximum. For this reason, we employ a stochastic algorithm based on Adaptive Simulated Annealing (ASA) combined with Newton’s method. Table 8.3 reports representative results of this reduction method applied to the estimation of the parameters of model (8.6). Note that Equation (8.6) may be generalized as follows: dpth;?

¼ kh ðmh

pth;? Þ dt

þ

½pth;? v

M X

shj dBtj ; for h ¼ 1; . . . ; M;

ð8:19Þ

j¼1

where the exponent n belongs to the interval [0,1.5] (Duffie 1996) (this restriction is needed for the existence and uniqueness of solutions). The discrete-time maximum-likelihood method does not involve particular difficulties with respect to the case we have dealt with (i.e. n 0.5). More precisely, the reduction method described by (8.12)
(8.18) can be easily adapted to this more general situation. Table 8.4 reports the results of a comparison (for the sake of simplicity, only two maturities are considered) between the estimates provided by model (8.6) and a simpler pureGaussian version (i.e. n  0) corresponding to a multivariate Vasicek model. Path generation using model (8.6) is carried out by means of the discrete version (8.8). The generation of a scenario with time frequency D
0 assigns the last observed data to /

TABLE 8.3 MLE of the Multivariate Model. Parameters for Some Representative Maturities Provided by the Multivariate Model (8.6) of the Orthogonal Components (pth;? of Equation (8.1)) of Italian Treasury Rates. The Estimate is Obtained from Daily Data of the 1999
/2005 Period by Means of the Techniques Described in Section 8.3.4.2 using Adaptive Simulated Annealing Maturity

k

m

6 months 36 months 60 months 360 months

4.3452 3.4351 3.4195 4.3450

0.4353 2.0429 2.8469 4.5369

s6 0.7356 0.0421 0.0081 0.0167

s36

s60

0.0421 0.2994 0.1871 0.0874

0.0081 0.1871 0.2455 0.1010

s360 0.0167 0.0874 0.1010 0.1993

SCENARIO-GENERATION METHODS

j

189

TABLE 8.4 Comparison of the Parameters of the Multivariate CIR Model (Equation (8.8)) and a Multivariate Vasicek (i.e. Pure Gaussian) Model, Corresponding to Equation (8.19) with n  0. For the Estimation We Resorted to the Discrete Maximum-Likelihood Method Applied to Daily Observations in the Period January 1999
/September 2005 of the pth;? Components of Two Maturities (60 and 120 Months, Indicated, Respectively, as 1 and 2). The Values of the t-statistics in the Multivariate CIR Model and the Larger Log-likelihood Indicate that Model (8.8) provides a Better Fit of the Original Data with Respect to the Pure-Gaussian Model Model Multivariate CIR Parameter k1 k2 m1 m2 s11 s12 s22 Log-likelihood

Multivariate Vasicek

Estimate

t-Stat.

Estimate

t-Stat.

1.6208 1.4175 1.6959 2.6422 0.5545 0.2788 0.3302 1001.925337

17.1201 15.0218 0.0917 0.5381 2.0263 2.0269 1.4483

1.4542 1.4243 1.6992 2.6557 0.6667 0.4079 0.5686 2206.122226

17.2518 16.4660 0.0950 0.5463 34.190 24.419 33.843

p1h for every h 1, . . . , M. Then, at each time step i, where 1 5i 5n 1, we generate independent random vectors zi from the multivariate normal distribution N(0, IM), and we set h ¼ pih piþ1

þ kh ðmh

pih ÞD

qffiffiffiffiffiffiffiffi þ pih DðC 1  zi Þh ;

ð8:20Þ

for i  1, . . . , n 1, where k, m and C are the MLE estimators. Figure 8.8 shows the results of a simulation, based on the multivariate CIR model described in this section, for a planning period of 48 months.

8.4

VALIDATION OF THE SIMULATED SCENARIOS

The stochastic models presented above allow the generation of a ‘realistic’ future term structure. However, we need to generate a (pretty long) temporal sequence of term structures starting from the present interest rates curve. As mentioned in Section 8.3, this requires control of the evolution in time of the simulated term structure in such a way that, for instance, its behavior is not too different from the behavior observed in the past. Besides control on the value of the simulated ECB with respect to a predefined lower bound as described in Section 8.3.1, we resort to two techniques to assess the reliability of the sequence of simulated term structures. The first method, which we classify as ‘local,’ ensures that, at each time step of the planning period, the simulated term structure is ‘compatible’ with the historical term structure. The second (which we classify as ‘global’) considers the whole term structure evolution, and ensures that the correlation among the increments of pth;? are close to the correlation of the increments of the historical fluctuations.

190 j CHAPTER 8 6.5

6.5 6

6

5.5

5.5

5

5

4.5

4.5

4 3.5 3 2.5 2

6 months 12 months 18 months 24 months 30 months 36 months 42 months 48 months

4 BCE BCEIN BOT 3 BOT 6 BOT 12 CTZ 24

0

5

10

15

20

25

30

35

BTP 36 BTP 60 BTP 120 BTP 180 BTP 360

40

45

50

3.5 3 2.5 2

0

Months

50

100

150 200 250 Maturity (in months)

300

350

FIGURE 8.8 Simulation of the evolution of interest rates based on the multivariate CIR model (left panel) and the corresponding term structures (right panel).

The local test is based upon the observation that the shape of the term structure does not vary widely over time. Obviously, there is some degree of variability. For instance, it is known that the term structure is usually an increasing function of the maturity, but, at times, it can be inverted for a few maturities. This is the fundamental assumption of Nelson and Siegel’s (1987) parsimonious model of the term structure. In the present case, since we are not interested in a functional representation of the term structure but in the interest rates at fixed maturities, we take as the indicator of the term structure shape the relative increment and the local convexity of the interest rates:

d ht ¼ rth rth 1 ;

cth ¼ r~th rth ;

ð8:21Þ

where r~th is the value at mh of the linear interpolation between rthþ1 and rth 1 :

r~ ht ¼ rth 1 þ

rthþ1 rth 1 mhþ1 mh 1

ðmh mh 1 Þ:

A positive convexity (cth > 0) means that the rate rth at maturity mh is below the line joining the rates rth 1 and rthþ1 at maturities mh 1 and mh, respectively, therefore the curvature opens upward. The case of cth < 0 can be interpreted along the same lines, but the curvature opens downward. At any time step, we accept the simulated interest rates if the corresponding values of dth and cth are not too different from the historical values. Briefly, we compute the historical mean and standard deviation (mhd , sdh ) and (mhc , sch ) of dth and cth , respectively, and then we check that

SCENARIO-GENERATION METHODS M 1 X h¼1

1 ðdth mhd Þ2  ðM 1Þ; h 2 ðsd Þ

M 1 X

1

h¼2

ðshc Þ

2

j

191

ð8:22Þ

ðcth mhc Þ2  ðM 2Þ:

The meaning of such tests is straightforward. If the simulated interest rates pass the tests, we are quite confident that their increments and the local convexity have statistical behaviour similar to the historical rates. The global test controls the correlation among the increments of pth;? , since it can be quite different from the correlation of the historical data. In order to avoid ‘pathological’ situations like anti-correlated increments or increments that are too correlated with each other, we compute the ‘one-norm’ of the matrix difference between the correlation matrix of the increments of the historical fluctuations and the correlation matrix of the simulated increments. The ‘one-norm’ is defined as follows: for each column, the sum of the absolute values of the elements in the different rows of that column is calculated. The ‘one-norm’ is the maximum of these sums: kShist i;j Si;j k1 ¼ sup i

M X

jShist i;j Si;j j:

ð8:23Þ

j¼1

We compare the result of this calculation with a predefined acceptance threshold. If the ‘one-norm’ is below the threshold, the simulated scenario is accepted, otherwise it is rejected and a new set of fluctuations is generated. Currently, the threshold is set equal to 0.05. Since the correlation is a number within the range [ 1,1], this simple mechanism guarantees that the covariance among the increments of the interest rates of different maturities in the simulated scenarios is pretty close to the historical covariance. Note how the local and global tests are complementary since the first test involves the shape of the term structure, whereas the second test concerns how the rates’ increments are correlated in time, that is, how the term structure evolves from one time step of the simulation to the next.

8.5

CONCLUSIONS AND FUTURE PERSPECTIVES

The management of public debt is of paramount importance for any country. Such an issue is especially important for European countries after the definition of compulsory rules by the Maastricht Treaty. Together with the Italian Ministry of Economy and Finance, we have studied the problem of finding an optimal strategy for the issuance of public debt securities. This turned out to be a stochastic optimal control problem where the evolution of the interest rates plays a crucial role. We have presented the techniques we employ to simulate the future behaviour of interest rates for a wide range of maturities (from 3 months up to 30 years). Since the planning period that we consider is pretty long (up to five years), most existing models

192 j CHAPTER 8

need to be modified in order to provide realistic scenarios. In particular, the evolution of the ‘leading’ rate (which we assume to be the European Central Bank’s official rate) is simulated as if it were a stochastic jump process. We are aware that there is a long tradition of studies that try to explain the links between prices, interest rates, monetary policy, output and inflation. Webber and Babbs (1997) were among the first to propose yieldcurve models able to capture some aspects of monetary policies. Piazzesi (2001) introduced a class of linear-quadratic jump-diffusion processes in order to develop an arbitrage-free time-series model of yields in continuous time that incorporates the central bank policy and estimates the model with U.S. interest rates and the Federal Reserve’s target rate. A general framework for these models has recently been presented by Das and Chacko (2002), where the authors introduce factors that influence the marginal productivity of capital, and thus the interest rates, in the economy. Their technique is general, since it applies to any multi-factor, exponential-affine term structure model with multiple Wiener and jump processes. From an empirical point of view, there is some evidence that important central banks, like the U.S. Federal Reserve, conduct a monetary policy (i.e. set the official rate) that is well described by the so-called Taylor’s rule (Taylor 1993). Basically, the rule states that the ‘real’ short-term interest rate (that is, the interest rate adjusted for inflation) should be determined according to three factors: (1) where the actual inflation is relative to the targeted level that the central bank wishes to achieve, (2) how far economic activity is above or below its ‘full employment’ level, and (3) what the level of the short-term interest rate is that would be consistent with full employment. Although Taylor’s rule appears to be more robust than more-complex rules with many other factors, it requires knowledge of inflation and the real Gross Domestic Product (GDP). The simulation of future inflation and GDP is far from easy, so, in some sense, the application of Taylor’s rule changes, but does not solve, the problem of generating meaningful scenarios for the evolution of the ECB official rate. The models proposed in the present chapter are fully integrated into the software prototype in use at the Ministry. The time required by the simulations (a few seconds on a personal computer) is such that we can afford on-line generation of the simulated scenarios even if the validation procedures described in Section 8.4 may require multiple executions before a scenario is accepted. Open problems and future analysis directions include the following: 



Modeling of the Primary Budget Surplus, Gross Domestic Product and Inflation in order to implement Taylor’s rule and possibly other models for the evolution of the European Central Bank’s official rate. Overcoming the assumption that interest rates are independent of the portfolio of existing securities and independent of the new securities issued every month. To limit the impact on the optimization problem, we should devise a description of these interactions that is compatible with the linear formulation of the problem.

SCENARIO-GENERATION METHODS

j

193

ACKNOWLEDGEMENTS We thank A. Amadori, D. Guerri and B. Piccoli for useful discussions. This work has been partially supported by a FIRB project of the Italian Ministry of Research.

REFERENCES Adamo, M., Amadori, A.L., Bernaschi, M., La Chioma, C., Margio, A., Piccoli, B., Sbaraglia, S., Uboldi, A. and Vergni, D., Optimal strategies for the issuances of public debt securites. Int. J. Theor. Appl. Finan., 2004, 7, 805
/822. Ait-Sahalia, Y., Maximum-likelihood estimation of discretely-sampled diffusions: A closed-form approach. Econometrica, 2002, 70, 223
/262. Balduzzi, P., Das, S., Foresi, S. and Sundaram, R., A simple approach to three factor affne models of the term structure. J. Fixed Income, 1996, 3, 43
/53. Baygun, B., Showers, J. and Cherpelis, G., Principles of Principal Components, 2000 (Salomon Smith Barney). Bergstrom, A.R., The history of continuous time econometric models. Economet. Theory, 1988, 4, 365
/383. Bibby, B.M. and Sorensen, M., Martingale estimation functions for discretely observed diffusion processes. Bernoulli, 1995, 1, 17
/39. Cox, J.C., Ingersoll, J.E. and Ross, S.A., A theory of the term structure of interest rates. Econometrica, 1985, 53, 385
/487. Dacunha-Castelle, D. and Florens-Zmirou, D., Estimation of the coefficients of a diffusion from discrete observations. Stochastics, 1986, 19, 263
/284. Das, S. and Chacko, G., Pricing interest rate derivatives: A general approach. Rev. Finan. Stud., 2002, 15, 195
/241. Duffie, D., Dynamic Asset Pricing Theory, 1996 (Princeton University Press: Pinceton, NJ). Duffie, D. and Kan, R., A yield-factor model of interest rates. Math. Finan., 1996, 6, 379
/406. Duffie, D., Pan, J. and Singleton, K., Transform analysis and asset pricing for affine jump-diffusions. Econometrica, 2000, 68, 1343
/1376. Jackson, D., The New National Accounts: An Introduction to the System of National Accounts 1993 and the European System of Accounts 1995, 2000 (Edward Elgar: Camberley, U.K.). James, J. and Webber, N., Interest Rate Modelling, 2000 (Wiley: New York). Litterman, R. and Scheinkman, J., Common factor affecting bond returns. J. Fixed Income, 1991, 1, 54
/61. Nelson, C.R. and Siegel, A.F., Parsimonious modeling of yield curves. J. Bus., 1987, 60, 473
/489. Papi, M., Term structure model with a central bank policy. In preparation, 2004. Piazzesi, M., An econometric model of the yield curve with macroeconomic jump effects. NBER Working Paper Series, 2001. Risbjerg, L. and Holmlund, A., Analytical framework for debt and risk management. Advances in Risk Management of Government Debt, 2005 (Organization for Economic Cooperation and Development). Taylor, J.B., Discretion versus policy rules in practice. Carnegie
/Rochester Conf. Ser. Public Policy, 1993, 39, 195
/214. Vasicek, O., An equilibrium characterization of the term structure. J. Finan. Econ., 1977, 5, 177
/188. Webber, N. and Babbs, S., Mathematics of Derivative Securities, 1997 (Cambridge University Press: Cambridge).

194 j CHAPTER 8

APPENDIX 8.A: ESTIMATION WITH DISCRETE OBSERVATIONS We deal with the case of a diffusion process Xt (i.e. the interest rate rt) that is observed at discrete times 0 ¼ t0 < t1 <    < tn , not necessarily equally spaced. If the transition density of Xt from y at time s to x at time t is p(x,t,y,s; u), where u is a collection of parameters to be estimated, we can resort to the Maximum Likelihood Estimator (MLE) b yn which maximizes the likelihood function Ln ðyÞ ¼

n Y

pðXti ; ti ; Xti 1 ; ti 1 ; yÞ;

ð8:A1Þ

i¼1

or, equivalently, the log-likelihood function

‘n ðyÞ ¼ logðLn ðyÞÞ ¼

n X

logðpðXti ; ti ; Xti 1 ; ti 1 ; yÞÞ:

ð8:A2Þ

i¼1

In the case of observations equally spaced in time, the consistency and asymptotic normality of y^n as n 0  can be proved (Dacunha-Castelle and Florens-Zmirou 1986). In general, the transition density of Xt is not available. In this case the classical alternative estimator is obtained by means of an approximation of the log-likelihood function for u based on continuous observations of Xt. Unfortunately, this approach has the undesirable property that the estimators are biased, unless p ¼ maxi jti ti 1 j is small. To overcome the difficulties due to the dependence of the parameters on p, different solutions have been proposed. One of the most efficient methods resorts to martingale estimating functions (Bibby and Sorensen 1995). This method is based on the construction of estimating equations having the following form: Gn ðyÞ ¼ 0;

ð8:A3Þ

where

Gn ðyÞ ¼

n X

gi 1 ðXti 1 ; yÞðXti IEðXti j Xti 1 ÞÞ;

ð8:A4Þ

i¼1

gi-1 being continuously differentiable in u, for i  1, . . . , n. Bibby and Sorensen (1995) proved, under technical conditions on gi-1, that an estimator that solves Equation (8.A3) exists with probability tending to one as n 0 , and this estimator is consistent and asymptotically normal. A simpler approach, which can be used when p is sufficiently small, is based on the Euler discretization of the diffusion equation associated with Xt. In this case, one can use the log-likelihood approach since the transition density can be easily computed. We discuss the application of this method to (8.5). Let

SCENARIO-GENERATION METHODS

rtiþ1 ¼ rti þ kðm rti ÞD þ s

qffiffiffiffiffiffiffi rti DZi ;

i ¼ 1; . . . ; n 1;

j

195

ð8:A5Þ

be the first-order discretization of Equation (8.5) on time interval [0, T], where D  T/n and with Zi the increment DBi of the Brownian motion between ti  iD and tiþ1 ¼ ði þ 1ÞD. Since these increments are independent N(0, D) distributed random variables, the transition density from y
0 to x during an interval of length D is   1 1 2 ðx kðm yÞDÞ ; pðx j y; yÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp 2ys2 D 2pys2 D

ð8:A6Þ

where y ¼ ðk; m; sÞ. Therefore, given a sequence of observations fri gi , an estimator of (8.32) is obtained by maximizing the log-likelihood function max y

n 1 X

ð8:A7Þ

logðpðrtiþ1 j rti ; yÞÞ:

i¼1

If one expands the function in (8.34) and sets equal to zero its partial derivatives with respect to ðk; m; s2 Þ, it is easy to show that the corresponding equations admit a unique solution that is a maximum point of the log-likelihood function. The following relations represent this ML estimator: m¼

s2 ¼

EF 2CD ; 2EB FD

k¼

D ; 2mB F

1 ðA þ k 2 m2 B þ k 2 C kmD þ kE k 2 mFÞ; n 1

ð8:A8Þ

ð8:A9Þ

TABLE 8.A1 Assessment of the Discrete Estimator (1D case). Results for a Simulation Evaluation of the Loglikelihood Estimator to the Univariate CIR Model. Using the True Values of the Parameters We Simulated 500 Sample Paths of Length 2610 Daily Observations Each. For Each Sample Path We Undertook Discrete MLE Estimation Via Euler Discretization. The Table Presents Summary Statistics of the Simulated Estimations. We Computed the Mean and Standard Error for the ^m ^; s ^), and Computed t-Statistics for the Difference between the Simulated Estimator (k; Parameter Estimate and the True Parameter. The Null Hypothesis Cannot be Rejected at the 5) Significance Level Parameter Statistics (N500)

k

m

s

True parameter Estimated parameter Standard error t-Statistics

0.8 0.79928 0.00081491 0.8835

3.24 3.2405 0.0011 0.4545

0.03 0.030009 0.00001936 0.4649

196 j CHAPTER 8 TABLE 8.A2 MLE of the Univariate CIR Model. Results for the Estimation of the Univariate CIR Model of pth;? of Equation (8.1). Estimation was Carried Out Using Discrete Maximum-likelihood (8.34). We Report the Value of the Log-likelihood (Log (L)), the Norm of its Gradient (|9Log (L)|) and the Maximum Eigenvalue (lL) of its Hessian Matrix at the Maximum Point Parameter

BOT3

BOT6

BOT12

CTZ

k m s log (L) |9log (L)| lL

5.4611 0.32712 0.78513 1866.541 1.2268 10-13 0.54044

4.0217 0.43815 0.76817 2174.4863 3.010310-15 1.1819

2.1634 0.72852 0.74801 1920.0013 6.349710-15 1.5844

1.1265 1.457 0.57573 1791.6797 2.273910-13 1.7826

BTP36

BTP60

BTP120

BTP180

BTP360

1.1488 2.0531 0.50163 1756.6086 1.4118 10-14 1.8906

1.4039 2.8686 0.4293 1743.7829 0 1.8071

1.7635 3.9144 0.33892 1822.044 9.0949 10-13 1.5539

1.9973 4.2957 0.31716 1839.9547 0 1.3116

2.4133 4.5355 0.30738 1840.362 1.50410-13 0.93247

k m s log (L) |9log (L)| lL

where A¼

n 1 X ðriþ1 ri Þ2 i¼0

D¼2

ri D

n 1 X riþ1 ri i¼0

ri

;

;

n 1 X 1 B¼D ; i¼0 ri

E ¼ 2ðrn r0 Þ;

C¼D

n 1 X

ri ;

i¼0

ð8:A10Þ

F ¼ 2Dðn 1Þ:

Here, ri is the interest rate at time ti. We report the results of this method in Table 8.A1. In Table 8.A2 the method is applied to each orthogonal component pth;? .

CHAPTER

9

Solving ALM Problems via Sequential Stochastic Programming FLORIAN HERZOG, GABRIEL DONDI, SIMON KEEL, LORENZ M. SCHUMANN and HANS P. GEERING CONTENTS 9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 9.2 Stochastic Programming Approximation of the Continuous State Dynamic Programming Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 199 9.2.1 Basic Problem and Description of the Algorithm . . . . . . . . . . . . . . 199 9.2.2 Scenario Approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201 9.2.3 Approximate of Dymanic Programming . . . . . . . . . . . . . . . . . . . . 202 9.3 Portfolio Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204 9.3.1 Asset Return Models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204 9.3.2 Portfolio Dynamics with Transaction Costs . . . . . . . . . . . . . . . . . . 205 9.3.3 Risk Measure and Objective Function . . . . . . . . . . . . . . . . . . . . . 206 9.3.4 Multi-Period Portfolio Optimization with Transaction Costs . . . . . . . 208 9.4 Case Study: Asset and Liability Management with Transaction Costs for a Swiss Fund . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209 9.4.1 Data Sets and Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209 9.4.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211 9.4.3 Factor Selection, Parameter Estimation and Asset Allocation Strategy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212 9.4.4 Results of the Historical Out-of-Sample Backtest . . . . . . . . . . . . . . 214 9.5 Conclusion and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218 Appendix 9.A: Additional Data for the Case Study . . . . . . . . . . . . . . . . . . . . 219 Appendix 9.B: Dynamic Programming Recursion for the Sample Approximation . . 220 197

198 j CHAPTER 9

9.1

I

INTRODUCTION

of real-world applications of asset liability management (ALM) with discrete-time models have emerged. Insurance companies and pension funds pioneered these applications, which include the Russell
Yasuada investment system (Carino and Ziemba 1998), the Towers
Perrin System (Mulvey 1995), the Siemens Austria Pension Fund (Ziemba 2003; Geyer et al. 2004), and Pioneer Investment guaranteed funds (Dempster et al. 2006). In each of the applications, the investment decisions are linked to liability choices, and the funds are maximized over time using multi-stage stochastic programming methods. Other examples of the use of stochastic programming to solve dynamic ALM problems are given by Dempster and Consigli (1998) and Dondi (2005). All authors propose stochastic programming as the most suitable solution framework for ALM problems. However, since most of the asset or liability models are dynamic stochastic models, the ALM problem is one of dynamic optimization which can be solved by applying the continuous state dynamic programming (DP) algorithm. In this chapter, we show that the DP algorithm can be approximated locally by stochastic programming (SP) methods. By using a sufficient number of scenarios, the difference between the exact solution and the approximation can be made arbitrarily small. The SP optimization is resolved for every time-step based on a new set of stochastic scenarios that is computed according to the latest conditional information. In this way, a feedback from the actual observed state is introduced which is not from the coarse scenario approximation from an earlier time-step. This procedure, often called rolling-horizon planning, is frequently used as heuristic, but we show here that by carefully posing and computing the SP optimization, the continuous state DP algorithm is being applied approximately. This DP approximation is applied to the problem of a fund that guarantees a minimal return on investments and faces transaction costs when investing. The situation resembles the problem faced by Swiss pension funds or German life insurance policies. The minimal return guarantee changes the fund problem from a pure asset allocation problem to an ALM problem, see Dempster et al. (2007). First, models of the asset returns and portfolio dynamics with transaction costs are introduced. Then we propose that the most suitable risk measure for such a situation is a shortfall risk measure that penalizes all possible scenarios for the future in which the minimal return is not achieved. The optimization problem is solved with the aim of maximizing the return above the guarantee over the planning horizon, while limiting the shortfall risk. The problem is tested in an eight year out-of-sample backtest with a quarterly trading frequency from the perspective of a Swiss Fund that invests domestically and in the EU markets and faces transaction costs. The chapter is organized as follows: in Section 9.2, the approximation of dynamic programming is introduced. In Section 9.3, we present the application of the stochastic programming approximation to a portfolio problem with transaction costs. Additionally, we introduce a dynamically coherent risk measure for asset liability situations. In Section 9.4, a case study is examined from the perspective of a Swiss fund that invests domestically in stocks, bonds and cash, as well as in the EU stock and bond markets. Section 9.5 concludes the chapter. N RECENT YEARS, A GROWING NUMBER

/

/

SOLVING ALM PROBLEMS VIA SEQUENTIAL STOCHASTIC PROGRAMMING

9.2

j

199

STOCHASTIC PROGRAMMING APPROXIMATION OF THE CONTINUOUS STATE DYNAMIC PROGRAMMING ALGORITHM

We discuss how to use dynamic stochastic programming as one possible approximation of dynamic programming (DP). It is well known that any dynamic optimization problem (DOP) in discrete-time may be solved in principle by employing the DP algorithm. However for situations with realistic assumptions, such as control constraints or nonGaussian white noise processes, it is very difficult to obtain closed-form solutions. In these cases it is necessary to resort to numerical solution methods. 9.2.1

Basic Problem and Description of the Algorithm

In this chapter, we approximate the continuous state dynamic programming method by discretizing not the state space but the possible outcomes of the white noise process. Starting from the current state value the DP approximation solves a SP problem over a finite horizon at each sampling time. The optimization procedure yields an optimal control sequence over the planning horizon but only the first control decision of the sequence is applied to the system and the rest are disregarded. At the next time-step, the calculation is repeated based on the new value of the state variable and over a shifted horizon, which leads to a receding horizon policy. The receding horizon policy means that we solve the multi-period decision problem always with the same number of periods to the horizon. Other authors have investigated the same idea in a portfolio optimization context, see Brennan et al. (1997), but based on numerical solutions of the Hamilton
Jacobi
Bellman PDE to solve the optimal control problem. A useful advantage of the present DP approximation is the capability to solve DP problems using established methods from stochastic programming and their known properties and algorithms. Other techniques, such as discrete DP solutions, often lead to computationally overwhelming tasks which often prevent their application. The technique proposed here however, solves the problem only for the current state and the approximate trajectory of the underlying dynamical system and thus avoids the curse of dimensionality. In Figure 9.1, the approximation algorithm is summarized; in Table 9.1 it is described in detail. /

/

Scenario Generation

Solving Stochastic Program u(t), us(t+1), .. ...u(t+T )s

Realization of ε(t) u(t)

Measure State (y(t))

FIGURE 9.1 Graphical description of the approximation algorithm.

y(t+1) = D(t, y(t)) + S(t, y(t)) ε(t)

200 j CHAPTER 9 TABLE 9.1 Algorithm to Compute the Stochastic Programming Approximation 1. 2. 3. 4. 5.

Based on the information at time t, determine y(t). Set an accuracy parameter h, set J ¼ 1, and define the number of scenarios per period. Start with a relatively low number of scenarios. Compute the scenario approximation of e(t) for t t, t  1, . . . , T  1 based on the number of samples per time period. Solve the multi-stage stochastic program as outlined in Proposition 9.1. In the case that jJ s ðt; yðtÞÞ  J j < Z stop and go to step 5. Otherwise set J :¼ J s ðt; yðtÞÞ and increase the number of scenarios to refine the approximation, see Section 9.4.2.1. Go to step 2. Apply only the first control decision uðtÞ and disregard all future control decisions for the next time-step. Until the optimization has reached its fixed end date (horizon) go to step 1, else stop.

A general DOP can be stated as ( " J ðt; yðtÞÞ : ¼ max E u2U

T 1 X

#) Lðt; yðtÞ; uðtÞÞ þ MðT ; yðT ÞÞ

t¼t

ð9:1Þ

s.t. yðt þ 1Þ ¼ Dðt; yðtÞ; uðtÞÞ þ Sðt; yðtÞ; uðtÞÞEðtÞ; where t  t, t  1, . . . , T  1, L( ×) and M( ×) are the strictly concave value functionals, D( ×) and S( ×) define the state dynamics and are assumed to be continuously differentiable functions, uðtÞ is a bounded control vector constrained to the convex set U and e(t) is a strictly covariance stationary white noise process. It is assumed that all functionals L( ×), M( ×), D( ×) and S( ×) are Lipschitz continuous and fulfil the necessary conditions for this dynamic optimization problem to be well defined and possesses a unique solution. For details refer to Bertsekas and Shreve (1978). Note that the white noise process is stationary, but the dynamics of y(t) are both state and time dependent since the functional Sðt; yðtÞ; uðtÞÞ depends on both y(t) and t. In order to obtain a feedback solution to the DOP problem (9.1), the dynamic programming (DP) algorithm is given by J ðT ; yðT ÞÞ ¼ MðT ; yðT ÞÞ; n o J ðt; yðtÞÞ ¼ max E½Lðt; yðtÞÞ þ J ðt þ 1; Dðt; yðtÞ; uðtÞÞ þ Sðt; yðtÞ; uðtÞÞ; EðtÞÞ : uðtÞ2U

ð9:2Þ This condition for optimality can be found in Bertsekas (1995, Chapter 1). Instead of solving the DP algorithm for J to yield the true stochastic dynamics, we locally approximate the stochastic dynamics by a finite number of scenarios at the current decision time and solve the problem repeatedly at each decision time-step. The standard DP procedure discretizes the state space for each dimension of y(t) and each time-step until the horizon T. Then, beginning at time T, the optimization problem of (9.2) is solved for each discretized state of y(t). Based on the optimal solution for each state discretization of y(t) the optimal value of J(T, y(T)) is known. The optimal control decision for time T  1 is solved by maximizing the backward recursion in (9.2). This procedure is repeated until we reach the current time t. In this way, the DP algorithm

SOLVING ALM PROBLEMS VIA SEQUENTIAL STOCHASTIC PROGRAMMING

j

201

solves (T  t  1)pk optimization problems, where p is the number of states, k is the number of discretizations for each state and T  t  1 is the number of remaining timesteps. For example, for a problem with five states, 10 discretizations and five remaining time-steps, the DP procedure involves 48 millon optimizations, which is already an enormous amount for this rather coarse discretization. This fact is often called the ‘curse of dimensionality’ since the number of computations increases exponentially with the number of states. 9.2.2

Scenario Approximations

If a regular grid is used to discretize the state space of (1), the optimization is computed at many points in the state space which are reached with low probability. A standard approach to overcome these drawbacks is to use a scenario approach and a sampling approximation of the true expectation. The difficulty with solving the DP problem directly is the computation of the recursion under realistic assumptions. Instead of computing the exact optimal control policy, we solve an approximate problem where we replace the expectation by the sample mean and the true dynamics by a finite number of representative scenarios. To calculate the scenario and sample mean, a number of samples has to be drawn from the underlying path probability distribution. This procedure is repeated at each time-step so that a new control decision is based on the current time and state. Instead of solving the DP problem for all possible states and time, the DP problem is approximated at the current state and time by an SP problem. At each time-step t, we replace the probability distribution of e(t) by k(t) scenarios, denoted by es(t). We denote by t  t, t  1, . . . , T  1 time in the scenario tree and thus in the SP problem. The ‘physical’ time is denoted by t. At time t  1 conditional on the scenario es(t), we generate k(t  1) scenarios for each previous scenario of es(t) as shown in Figure 9.2. Since we assume that the white noise process is stationary, at each node e(t) is replaced by the same scenario values. The system dynamics however are different at each node, since both D( ×) and S(×) are time and state dependent. The scenarios are defined as the set S Number Branches for Each Node

Scenario Tree

24

5

2

Time Steps Time 1

Time 2

Time 3

Time 4

FIGURE 9.2 Graphical description of a scenario tree.

202 j CHAPTER 9

that represents a reasonable description of the future uncertainties. A scenario s
S describes a unique path through consecutive nodes of the scenario tree, as depicted in Figure 9.2. In this way, we generate a tree of scenarios that grows exponentially with the QT time horizon and the number of scenarios is given by N ¼ i¼t kðiÞ. By this procedure we generate an irregular grid of the state space along a highly probable evolution of the system. The optimization is then computed for the approximate dynamics in terms of the sample expectation. For the approximate problem we still effectively execute the DP algorithm, but only for the approximate paths of the state dynamics and the current state and time. 9.2.3

Approximate of Dymanic Programming

When we replace the ‘true’ stochastic dynamics of the state equation by its sample approximation, we need to compute the sample mean of the objective function instead of its expectation. The objective function at time t on one scenario s with feedback mapping (policy) Ps ðtÞ :¼ ½ps ðtÞ; ps ðt þ 1Þ; . . . ; ps ðT  1Þ is given by

V s ðt; y s ðtÞ; Ps ðtÞÞ ¼

T 1 X

Lði; y s ðiÞ; ps ðiÞÞ þ MðT ; y s ðT ÞÞ:

ð9:3Þ

i¼t

The control decisions Ps(t) define a feedback mapping, since for each scenario s and time t a predetermined control decision based on a feedback rule ps(t):  u(t, ys(t)) is used. The sample mean of the objective function is given by N i 1X h V s ðt; y s ðtÞ; Ps ðtÞÞ; E^ V s ðt; y s ðtÞ; Ps ðtÞÞ ¼ S s¼1

ð9:4Þ

P ^ ¼ ð1=SÞ Ss¼1 ½ denotes the sample mean. Using (9.4) we define the sample where E½ approximation of the dynamic optimization problem as J^s ðt; yðtÞÞ :¼ max s

P ðtÞ2U

n h io E^ V s ðt; y s ðtÞ; Ps ðtÞÞ

ð9:5Þ

s.t. y s ðt þ 1Þ ¼ Dðt; y s ðtÞ; ps ðtÞÞ þ Sðt; y s ðtÞ; ps ðtÞÞEs ðtÞ; where J^s ðt; yðtÞÞ is the value of the objective function at time t under scenario s and we must impose non-anticipativity constraints, i.e. y(t)s  y(t)s’ when s and s? have the same past until time t. At the root of the scenario tree J^s ð Þ and ps(t) are the same for all scenarios (s  1, . . . , N?). Theorem 9.1: The sample approximation of the optimization problem given in (9.5) can be recursively computed by the following dynamic program:

SOLVING ALM PROBLEMS VIA SEQUENTIAL STOCHASTIC PROGRAMMING

n h J^s ðt; y s ðtÞÞ ¼ max E^ Lðt; y s ðtÞ; us ðtÞÞ þ J^s ðt þ 1; Dðt; y s ðtÞ; uðtÞ2U io us ðtÞÞ þ Sðt; y s ðtÞ; us ðtÞÞEs ðtÞÞÞ

j

203

ð9:6Þ

with terminal condition J^s ðT y s ðT ÞÞ ¼ MðT ; y s ðT ÞÞ. The proof of Theorem 9.1 is given in Appendix 9.B. Note that using (9.6) to solve (9.5), we use a backward recursion that automatically takes into account the non-anticipative constraints which prevent exact knowledge of future scenarios. Proposition 9.1: The dynamic programming formulation of the approximated dynamic optimization problem can be written as a multi-stage stochastic program. This proposition follows directly from Louveaux and Birge (1997, Chapter 3, p. 128). In Table 9.1 we state the algorithm to compute the stochastic programming approximation of the dynamic programming approach. By starting with a low number of scenarios, we ensure that the multi-stage stochastic program is solved rather quickly. The algorithm cycles between step 2 and step 4 until the desired accuracy has been reached. The relation of the approximation algorithm and the true problem defined in (9.1) is given in Proposition 9.2. Proposition 9.2: Under the assumptions of Section 9.2.1, the sample approximation of the objective function J^s ðt; yðtÞÞ defined in (9.5) converges with probability 1 to the true objective function J(t, y(t)) as defined in (9.1) for N 0 . Especially J^s ðt; yðtÞÞ converges with probability 9.1 to J(t, y(t)) for N 0 . Proof: Given a predetermined feedback mapping Ps(t) as defined above and using the Tchebychev inequality (Casella and Berger 2002) the following holds:

 ^ ðt; y s ðtÞ; Ps ðtÞÞ  E½V ðt; yðtÞ; PðtÞÞ < Z ¼ 1; lim P E½V

8Z > 0;

N !1

ð9:7Þ

where t  t, t  1, . . . , T  1. We know that exchanging the expectation with the sample ^ ðt; y s ðtÞ; Ps ðtÞÞ mean has a negligible effect with arbitrarily large probability, since E½V converges to E½V ðt; yðtÞ; PðtÞÞ with probability 1 and using J^s ðt; y s ðtÞÞ ¼ max s

P ðtÞ2U

n h io E^ V s ðt; y s ðtÞ; Ps ðtÞÞ ;

it follows under suitable assumptions that

 E½J^ðt; y s ðtÞÞ  E½J ðt; yðtÞÞ < Z ¼ 1; lim P ^

N!1

8Z > 0:

ð9:8Þ I

204 j CHAPTER 9

The proof of Proposition 9.2 holds only under the restrictive assumptions for the problem stated in Section 9.2.1. The more general case is discussed in detail in Ro¨misch (2003) and Dempster (2004). The result of Proposition 9.2 states that the approximate objective function converges in probability to the true value of the objective function. However this does not imply that the control law computed by the approximation converges in probability to the true control law. The approximation by stochastic programming techniques determines only the control law for the scenarios used. For other values of the state variables, which are not covered by the scenario approximation, the control law is not defined. Since we only apply the first control decision for the fixed (measured) state y(t) the issue of feedback for uncovered state variables may not constitute a problem in the absence of nonlinearities. By using the approximation procedure at every time-step, however, the control decisions are always based on current information, but remain scenario dependent. A similar analysis for linear stochastic systems with quadratic performance criteria can be found in Batina et al. (2002), however without the explicit connection to stochastic programming. The scenario approximation does not depend on the dimension of the state variables but on the number of scenarios used. The algorithm’s complexity is thus independent of the state space dimension. However, to obtain results with a desired accuracy we need a sufficiently large number of scenarios. By solving the stochastic programming problem at every time-step we introduce feedback into our system. This method requires the solution of T  t  1 stochastic programming problems. The approach is very suitable for historical backtesting, but is less suitable for simulation studies, since for each time-step in the simulation a stochastic program must be solved. Furthermore, this approach is limited by the exponential growth of the scenario tree. If a very large number of scenarios is needed, it becomes very slow and computationally intractable. The convergence speed of the proposed method relies on the convergence speed of the stochastic program and the scenario generation. As shown in Koivu (2005), the scenario generation method determines the convergence speed and the accuracy of the solution. For this reason we use the method proposed in Koivu (2005) for scenario generation, since it outperforms standard Monte Carlo techniques (see Section 9.4.2.1).

9.3

PORTFOLIO OPTIMIZATION

In this section, the general asset return statistical model and the portfolio model with transaction costs are given. Further, the objective function and the problem of portfolio optimization under transaction costs is stated. 9.3.1

Asset Return Models

The returns of assets (or asset classes) in which we are able to invest are described by rðt þ 1Þ ¼ GxðtÞ þ g þ Er ðtÞ; T

ð9:9Þ

where rðtÞ ¼ ðr1 ðtÞ; r2 ðtÞ; . . . ; rn ðtÞÞ 2 Rn is the vector of asset returns, Er ðtÞ 2 Rn is a white noise process with E½Er ðtÞ ¼ 0 and E½Er ðtÞErT ðtÞ :¼ SðtÞ 2 Rnn is the covariance

SOLVING ALM PROBLEMS VIA SEQUENTIAL STOCHASTIC PROGRAMMING

j

205

matrix, GxðtÞ þ g 2 Rn is the local expected return, xðtÞ 2 Rm is the vector of factors, G 2 Rnm is the factor loading matrix, and g 2 Rn is a constant vector. We assume that the conditional expectation is time-varying and stochastic, since G x(t)  g is a function of the factor levels x(t) which themselves are governed by a stochastic process. The white noise process er(t) is assumed to be strictly covariance stationary. The prices of the risky assets evolve according to   Pi ðt þ 1Þ ¼ Pi ðtÞ 1 þ ri ðtÞ ;

Pi ð0Þ ¼ pi0 > 0;

ð9:10Þ

where P(t)  (P1(t), P2(t), . . . , Pn(t)) denotes the prices of the risky assets. A locally riskfree bank account with interest rate r0(t, x(t)) is given as   P0 ðt þ 1Þ ¼ P0 ðtÞ 1 þ F0 xðtÞ þ f0 ;

P0 ð0Þ ¼ p00 > 0;

ð9:11Þ

where P0(t) denotes the bank account. The interest rate of the bank account, described by (9.11), is modelled by r0 ðtÞ ¼ F0 xðtÞ þ f0 ;

ð9:12Þ

where F0 2 R1m and f0 2 R. The factor process affecting the expected return of the risky assets and the interest rate of the bank account is described by the following linear stochastic process difference equation xðt þ 1Þ ¼ AxðtÞ þ a þ nEx ðtÞ;

ð9:13Þ

where A 2 Rmm , a 2 Rm , n 2 Rmm and Ex ðtÞ 2 Rm is a strictly covariance stationary white noise process. The white noise process of the risky asset dynamics er(t) is not restricted to have a Gaussian distribution. We also assume that er(t) and ex(t) are correlated. The stochastic process of the asset returns has a Markov structure and therefore we can apply DP techniques to solve the corresponding portfolio optimization problem. 9.3.2

Portfolio Dynamics with Transaction Costs

For the case of transaction costs, we limit our description of the wealth dynamics to linear transaction costs and use the scenario approach to describe multi-period asset prices. Many different formulations of multi-period investment problems can be found in the literature. Here, we adopt the basic model formulation presented in Mulvey and Shetty (2004). The portfolio optimization horizon consists of T time-steps represented by t  {1, 2, . . . , T }. At every time-step t, the investors are able to make a decision regarding their investments and face inflows and outflows from and to their portfolio. The investment classes belong to the set I  {1, 2, . . . , n}. Let zis ðtÞ be the amount of wealth invested in instrument i at the beginning of the timestep t under scenario s. The units we use are the investor’s home currency (e.g. Swiss francs). Foreign assets, hedged or un-hedged against exchange rate fluctuations, are also

206 j CHAPTER 9

denoted in the portfolio’s home currency. At time t the total wealth of the portfolio is W s ðtÞ ¼

n X

zis ðtÞ

8s 2 S;

ð9:14Þ

i¼1

where W s(t) denotes the total wealth under scenario s. Given the returns of each investment class, the asset values at the end of the time period are zis ðtÞð1 þ ris ðtÞÞ ¼ z~is ðtÞ

8s 2 S;

8i 2 I;

ð9:15Þ

where ris ðtÞ is the return of investment class i at time t under scenario s. The returns are obtained from the scenario generation system. Therefore, z~is ðtÞ is the ith asset value at the end of the time period t under scenario s. The sales or purchases of assets occur at the beginning of the time period, where dis ðtÞ  0 denotes the amount of asset i sold at time t under scenario s, and pis ðtÞ  0 denotes the purchase of asset i at time t under scenario s. The asset balance equation for each asset is zis ðtÞ ¼ z~is ðt  1Þ þ pis ðtÞð1  di Þ  dis ðtÞ

8s 2 S;

8i 2 I n f1g;

ð9:16Þ

where di is the proportional (linear) transaction cost of asset i. We make the assumption that the transaction costs are not a function of time, but depend only on the investment class involved. We treat the cash component of our investments as a special asset. The balance equation for cash is z1s ðtÞ

¼

z~1s ðt

 1Þ þ

n X

dis ðtÞð1

 di Þ 

n X

i¼2

pis ðtÞ þ qs ðtÞ

8s 2 S;

ð9:17Þ

i¼2

where z1s ðtÞ is the cash account at time t under scenario s and qs(t) is the inflow or outflow of funds at time t under scenario s, respectively. All of the variables in Equations (9.14)
(9.17) are dependent on the actual scenario s. These equations could be decomposed into subproblems for each scenario in which we anticipate that this scenario will evolve. To model reality, we must, however, impose nonanticipativity constraints. All scenarios which inherit the same past up to a certain time period must evoke the same decisions in that time period, otherwise the non0 anticipativity requirement would be violated. So zis ðtÞ ¼ zis ðtÞ when s and s? have same past until time t. /

9.3.3

Risk Measure and Objective Function

We introduce a linear risk measure that is well suited for problems with assets and liabilities. Liabilities can be explicit payments promised at future dates, as well as capital guarantees (promises) to investors or investment goals. We define our risk measure as a penalty function for net-wealth, i.e. wealth minus liabilities. We want to penalize small ‘non-achievement’ of the goal differently from large

SOLVING ALM PROBLEMS VIA SEQUENTIAL STOCHASTIC PROGRAMMING

j

207

‘non-achievement.’ Therefore, the penalty function should have an increasing slope with increasing ‘non-achievement.’ The risk of the portfolio is measured as one-sided downside risk based on non-achievement of the goals. As penalty function Pf (×) we choose the expectation of a piecewise linear function of the net wealth as shown in Figure 9.3. The penalty function is convex, but not a coherent risk measure in the sense of Artzner et al. (1999) However, it fulfills the modified properties for coherent risk measures defined by Ziemba and Rockafellar (2000). Furthermore, we can formulate a backward recursion to compute the risk measure and, thus, it is dynamically coherent in the sense of Riedel (2004). The same approach is discussed in detail by Dondi et al. (2007) where this method is applied to the management of a Swiss pension fund. Furthermore, for the case of multi-period capital guarantees, a suitable risk measure that is linear is given by G¼

" N T X X s¼1

#  s  Pf W ðtÞ  GðtÞ ;

ð9:18Þ

t¼t

where Pf denotes the piecewise linear penalty function, G(t) denotes the capital guarantee at time t and W(t) the portfolio value at time t. By multi-period capital guarantee we mean a capital guarantee not only for the final period but for all intermediate periods. This risk measure is convex and fulfills the properties of a dynamic risk measure. Standard coherent risk measures (CVaR, maximum loss) or traditional risk measures (utility functions, variance, VaR) can be used in ALM situations, when applied to the fund’s net wealth, e.g. the sum of the assets minus all the present value of the remaining liabilities. For example, CVaR penalizes linearly all events which are below the VaR limit for a given confidence level. The inherent VaR limit is a result of the CVaR optimization, see Rockafellar and Uryasev (2000, 2002). The VaR limit therefore depends on the confidence level chosen and the shape of the distribution. The VaR limit (quantile) may be a negative number, i.e. a negative net wealth may result. For a pension fund we do not only want to penalize scenarios that are smaller than a given quantile, but all scenarios

Negative Net Wealth

Penality Function

Positive Net Wealth

Net Wealth

FIGURE 9.3 Depiction of the penalty function.

208 j CHAPTER 9

where the net wealth is non-positive. Similarly, traditional risk measures do not measure the shortfall, but the inherent ‘randomness’ of the investment policy. We have therefore chosen the approach described in the previous paragraph. The objective of our ALM problem is to maximize the expected wealth at the end of the planning horizon, while keeping the penalty function under a certain level. Therefore, we propose the following objective function

max

( " N X

s

W ðT Þ þ l

T X t¼t

s¼1

 

Pf t; W s ðtÞ  GðtÞ F ðtÞ

#) ð9:19Þ

;

where t  t, t  1, . . . , T, and l B 0 denotes the coefficient of risk aversion. 9.3.4

Multi-Period Portfolio Optimization with Transaction Costs

For the case of portfolio optimization problems with long-term goals and transaction costs we use the approach presented in Section 9.2. The optimization problem for the minimal guaranteed return fund is

max

( " S X

pis ðtÞ;dis ðtÞ

s

W ðtÞ ¼ zis ðtÞð1 þ ris ðtÞÞ ¼ zis ðtÞ ¼

n X

s

W ðT Þ þ l

t¼t

s¼1

zis ðtÞ

i¼1 z~is ðtÞ z~is ðt 

T X

 

Pf t; W ðtÞ  GðtÞ F ðtÞ

#) ;

8s 2 S; 8s 2 S;

1Þ þ

z1s ðtÞ ¼ z~1s ðt  1Þ þ

8i 2 I;

pis ðtÞð1  di Þ  dis ðtÞ 8s 2 S; 8i n n X X dis ðtÞð1  di Þ  pis ðtÞ þ qs ðtÞ i¼2 i¼2

2 Inf1g;

ð9:20Þ

8s 2 S;

where t  t, t  1, . . . , T, l B 0, denotes coefficient of risk aversion, and initial conditions zi(t), i  1, . . . , n are given. The optimization problem is given directly as the SP problem that needs to be solved at every time-step. Furthermore, we may impose constraints for short positions or leveraging and we may limit the position in a specific asset (or asset class) by imposing the constraints ðminÞ

ui ðminÞ



zis ðtÞ W s ðtÞ

ðmaxÞ

 ui

ðmaxÞ

;

where ui is the minimum allowed fraction and ui is the maximum allowed fraction of wealth invested in asset i respectively. Similar constraints can be introduced to enforce minimum and maximum investments into certain sets of assets, e.g. international investments, or stocks. This kind of constraint can be formulated as linear and thus, the

SOLVING ALM PROBLEMS VIA SEQUENTIAL STOCHASTIC PROGRAMMING

j

209

optimization given in (9.20) is still a large-scale LP. Many specialized algorithms exist to solve this special form of an LP, such as the L-shaped algorithm, see e.g. Birge et al. (1994). The portfolio optimization problem is solved by employing the algorithm outlined in Section 9.2. At every time-step, we generate the scenario approximation of the return dynamics used for the specific investment universe based on the current information at time t. Then, we solve the optimization problem given by (9.20) with all of the corresponding constraints. Next we then check the accuracy and refine the scenario tree. When we reach the predetermined accuracy, we terminate the tree generation and optimization and apply the investment decisions pi(t) and pi(t). The procedure then moves one time-step ahead.

9.4

CASE STUDY: ASSET AND LIABILITY MANAGEMENT WITH TRANSACTION COSTS FOR A SWISS FUND

In this case study, we take the view of an investment fund that resides in Switzerland which invests domestically and abroad. The fund is assumed to be large and we therefore cannot neglect the market impacts of its trading activities which result in trading costs. The problem of portfolio optimization is solved with the framework presented in Section 9.3.2. Moreover, we assume that the fund gives a capital (performance) guarantee which introduces a liability. The situation resembles the situation of a Swiss pension fund and, thus, we impose similar restrictions on the case study. Designing funds with performance guarantees is also discussed in Dempster et al. (2006, 2007). The critical connection between the assets and liabilities is modelled through the capital guarantee GðtÞ. In the case of fixed discount rates, such as in the Swiss or German case, the capital guarantee increases with the discount rate. For the case in which the capital guarantee is linked to the current term structure of interest rates, the capital guarantee would increase along the term structure. The guarantee is not necessarily deterministic, it could be stochastic, e.g. the guarantee could increase with a LIBOR rate that changes in the future. Then the guarantee must be part of the stochastic scenario generation. In this case study, we assume that the discount rate is fixed to 4) and the nominal liabilities increases accordingly. Therefore, we do not include the capital guarantee in the scenario generation. Other approaches that feature a two-step method to calculate the optimal ALM are described in Mulvey et al. (2000), where the liabilities are described by a detailed interest rate model. A three factor yield curve model is also used in Dempster et al. (2006, 2007). Classical techniques are the immunization of liabilities by bonds, as described in Fabozzi (2005). A detailed overview of different approaches to ALM modelling is given in Dondi et al. (2007) and Zenios and Ziemba (2007). 9.4.1

Data Sets and Data Analysis

The data sets consists of the Datastream (DS) Swiss total stock market index, the DS Swiss government benchmark bond index, the DS European Union (EU) total stock market, and the DS EU government benchmark bond index. For the money market account, we use the 3-month LIBOR (SNB) interest rate. The data set starts on 1 January 1988 and finishes on 1 January 2005 with quarterly frequency. The two international indices are used in two

210 j CHAPTER 9

different ways in our case study. In the first the indices are simply recalculated in Swiss francs (CHF) and the in other the currency risk is eliminated by completely hedging the currency risk. The risk-return profile is changed in the latter case since hedging introduces costs that reduce the performance but eliminate the currency risk. The hedging costs are computed on the basis of the difference of the 3-month forward rates between the Swiss franc and euro. Before 1999, we use the difference between forward rates of the Swiss franc and German mark as an approximation for the Euro
Swiss France hedging costs. Also, before 1999 the EU stock and bond indices are calculated in German marks. We simply substitute the euro by German mark (with euro reference conversion of 1.95583), since the German mark was the major currency of the euro predecessor the ECU (the German mark and linked currencies such as the Dutch gilder made up more than 50) of the ECU). The correlation of the ECU and the German mark were extremely high, especially after 1996 where the out-of-sample backtesting starts. This substitution is of course a simplification, but does not change the return distributions significantly. In Figure 9.4, the return histograms for the Swiss stock market, the Swiss bond market, the EU stock market in CHF, and the EU stock market hedged are shown. The figure shows the histograms and the best fits of a normal distribution. Except for the bond market index, we can clearly reject the assumption that the stock market data are normally distributed. When we fit a multivariate student-t distribution on a rolling basis to the stock market data, we get degrees of freedom between 6.7 and 9.6, which indicate a very clear deviation from normality. These results are supported by two tests for normality which reject the assumption of normality at 5) confidence level. The two tests are the Jarque
Bera test /

/

Number of Observations

Number of Observations

15 10 5 0 –0.4

–0.2

0.2

10 5 0 –0.1

0.4

–0.05

15 10 5

–0.2

0 Return in %

0

0.05

0.1

Return in %

Number of Observations

Number of Observations

0 Return in %

15

Normal dist. fit EU stock market returns in CHF 15

20

0 –0.4

Normal dist. fit Swiss bond market returns

Normal dist. fit Swiss stock market returns 20

20

0.2

0.4

Normal dist. fit EU stock market returns hedged

10

5

0 –0.4

–0.2

0 Return in %

0.2

0.4

FIGURE 9.4 Histogram of the quarterly returns for different assets of the Swiss case study.

SOLVING ALM PROBLEMS VIA SEQUENTIAL STOCHASTIC PROGRAMMING

j

211

(J
B) and the Lilliefors test (LF) (Alexander 2001, Chapter 10). For these reasons, we model the returns of the stock market data by a non-normal distribution. The results of the normality tests for the out-of-sample period are given in Table 9.4, Section 9.4.4. We obtain similar results for the in-sample time period from 1 January 1988 to 1 June 1996. /

9.4.2

Implementation

The implementation of the out-of-sample test for the portfolio allocation method consists of three main steps: the scenario generation, the parameter estimation and the computation of the asset allocation. By using the methods discussed in Section 9.3.2 one crucial step of the implementation is the generation of scenarios. The scenarios describe the future stochastic evolution of the asset returns and must reasonably well approximate the underlying stochastic model. 9.4.2.1. Scenario generation The most common techniques to generate scenarios for multistage stochastic programs are discussed in Dupacova et al. (2000). Among the most important methods of scenario generation are moment matching (Hoyland et al. 2003), importance sampling (Dempster and Thompson 1999, Dempster 2004) or discretizations via integration quadratures (Pennanen 2005, Pennanen and Koivu 2005). We use the method of discretization via integration quadratures, because we believe that this method is superior to Monte Carlo methods, especially for high dimensional problems. Furthermore, numerical tests validate the stability of the optimization results, as shown in Pennanen and Koivu (2005). The method is used by approximating the white noise process at every stage of the dynamic model. Since we have assumed that the white noise process is stationary and identically distributed, we can use the same scenario generation method for er(t) and ex(t) for each time-step and scenario. The resulting scenarios of the dynamic stochastic evolution of the system are different for each scenario, since the asset return evolution depends on the evolution of the factors and the correlation of asset return dynamics and factors. The scenarios for the asset and portfolio evolution therefore become stochastic and dynamic. For the generation of the low discrepancy sequences, which are essential for the discretization we use the Sobol sequence, see Bratley and Fox (1988). We discuss the implementation of scenario generation for the student-t distribution which we assume for the white noise process of the asset returns. The multivariate student-t distribution possesses the following parameters: mean vector m, degrees of freedom n and diffusion matrix S. The algorithm to compute an s-sample from a multivariate student-t distribution is based on the following result (Glasserman 2004, Chapter 9, p. 510):

X  N ðm; SÞ;

z  w2n ;

Y ¼

pffiffiffi X n pffiffiffi  tnn ; z

where N denotes the normal distribution in Rn , w2n the standard chi-square distribution with n degrees of freedom, and tnn the student-t distribution in Rn with n degrees of freedom. The covariance of the student-t distribution is given by n=ðn  2ÞS and exists

212 j CHAPTER 9 TABLE 9.2 Scenario Generation for a Multivariate Student-t Distribution 1. Use the Sobol sequence to generate an s-sample from the marginally independent uniform distribution in the unit cube [0, 1]n1 distributions, denoted by S
[0, 1](n  1)  s. 2. Use the tables of the inverse standard normal distribution to transform the uniformly distributed realizations (Si, i  1 ,. . ., n) to standard normal distributed random variables (e xi i ¼ 1; . . . ; n), i.e. e xi ð jÞ ¼ F1 ðSi ð jÞÞ, i  1 , . . . , n, j 1, . . . , s. The realizations of the n normal distribution are summarized in the vector e xð jÞ ¼ ½e x1 ð jÞ; . . . ; e xn ð jÞ , j  1 , . . . , s. Use tables of the inverse of the chi-square distribution to transform Sn1(j) to chi-square distributed realizations which we denote by f(j), j  1 , . . . , s. e :¼ covðe 3. Compute the covariance matrix S xÞ, the covariance matrix of the realizations of the normal e 1 e xð jÞ, distribution of step 2, and calculate the normalized sequence of standard random variables xð jÞ ¼ S j 1 , . . . , s. In this way, we ensure that xðsÞ possessespunit variance. pffiffiffiffiffiffiffiffiffiffi ffiffiffi 4. Calculate the realizations of the white noise by Eð jÞ ¼ nðm þ sxð jÞÞ= fð jÞ, j  1 , . . . , s where S ¼ ssT , m and n are the parameters of the multivariate student-t distribution.

only for n  2. The scenario generation algorithm for the student-t distribution is given in Table 9.2. With similar algorithms for scenario generation, any kind of normal (variance) mixture distributions as defined in McNeil et al. (2005, Chapter 3, p. 78) can be approximated, as long as the method of inverses (Glasserman 2004, Chapter 2, p. 54) can be applied. Distributions such as the generalized hyperbolic distribution which belong to the family of normal (variance) mixture distributions can be expressed as a function of the multivariate normal distribution and a mixing univariate distribution. The normal distribution can again be simulated by generating n independent realizations of the uniform [0, 1] distribution and using the tables of the inverse univariate standard normal distribution. This is possible since we can generate the multivariate normal distribution from its univariate marginal distributions and the Cholesky factorization of its covariance matrix. For the univariate mixing distribution also realizations of the uniform [0, 1] distribution are generated and realizations of the mixing distribution are calculated from tables of the inverse. The resulting realizations of the non-normal distributions are generated by using the functional relationship of the normal (variance) mixture distributions. 9.4.3

Factor Selection, Parameter Estimation and Asset Allocation Strategy

The factor selection determines which of the factors best explain the expected returns of the risky assets. Factor selection is used for the regression problem between the expectation of the risky asset returns Gx(t)  g and the factors x(t). Often, this selection is predetermined using literature recommendations or economic logic. However, when we use a very large set of factors, it is difficult to decide which factors explain the expected returns best. Moreover, by using a predefined set of factors a bias is introduced into the out-of-sample test, since from knowing the history we include factors where it is known that they have worked as predictors. In order to solve this problem, we employ a heuristic as described in Givens and Hoeting (2005, p. 55) known as stepwise regression. When we want to choose the best possible subset of factors, we face a combinatorial number of possible subsets. For this reason we use a ‘greedy’ strategy that reduces the number of factors by the factor with the lowest impact. Given m factors, the heuristic creates only m subsets and we choose the best subset by a so-called information criterion.

SOLVING ALM PROBLEMS VIA SEQUENTIAL STOCHASTIC PROGRAMMING

j

213

The information criterion is a trade-off between the number of regressors and the regression quality. For factor selection we either use the modified Akaike criterion or the Schwartz
Bayesian criterion. The information criterion here is used to select the best regression model for the expected return, but is not applied to select the best distribution. This factor selection procedure is used for all risky assets independently. Different expected returns of the risky assets are regressed on different sets of factors. The factor selection is recomputed every 12 months, i.e. the factors are selected at the beginning of every year. In this way, we remove of the factor selection bias and the model selects only the factors that have worked in the past without any information about the future. This heuristic and the two information criteria are discussed in detail in Illien (2005). The estimation of parameters for the risky return is computed on a rolling basis with the last 8 years of data used. See Table 9.A1 in Appendix 9.A for a list of potential factors. The factor selection determines which time series of factors are used to predict the expected returns of the four risky assets. We assume that the fund faces different transaction costs for domestic and international assets. The costs (due to market impact and international brokerage cost) for the domestic market are 1) for stocks and 0.5) for bonds. For the European stock market, the transaction costs are assumed to be 2) and for the bond market 1). The transaction costs for European assets are independent of the hedging, since we calculate the hedging cost as part of the realized returns in Swiss francs. The asset allocation decisions are calculated with the optimization algorithm proposed in (9.20). We assume that the fund possess a two-year moving investment horizon and we use a tree structure with 50, 20 and 5 branches which results in 5000 scenarios which we denote by 5000 (50,20,5). For the first branching, we use a one quarter time-step, for the second branching we use two quarters, and for the third branching we use five quarters. The algorithm to approximate (locally) the DP algorithm given in Section 9.2 is implemented by first computing 500 (10,10,5), 1000 (20,10,5), 3000 (20,15,10), 4000 (40,20,5) and 5000 (50,20,5) scenarios. The relative error between using 4000 and 5000 scenarios was smaller than 1) (measured by the objective function value obtained with 5000 scenarios). This test was done at the first time-step of the out-of-sample test and repeated every 12 quarters. In all tests, the difference was smaller than the 1). The constraints for the optimization are similar to the constraints which Swiss pension funds face. In Table 9.3 the maximum limits for investments in the different asset classes are given. /

TABLE 9.3 Investment Constraints Swiss stock market Swiss bond market EU stock market EU bond market EU stock market in CHF (hedged) EU bond market in CHF (hedged) All international assets All stock market investments

50) 100) 25) 70) 40) 70) 70) 50)

214 j CHAPTER 9

As mentioned before, we assume that the fund gives the same capital guarantee of 4) as a Swiss pension fund (until 2002). Since 2002, the minimum guarantee has been reduced to the region of 2.5) and, therefore, we calculate a second portfolio with minimal guarantee of 2.5). The capital guarantee function is given as G(q) 100(1  ru4)q where r is either 4) or 2.5) and q is the number of quarters in the backtesting. When we solve the optimization problem at every time-step, the current asset allocation is taken as the initial asset allocation. In this way, the transaction costs for every rebalancing of the portfolios are correctly considered. The risk measure used for this case study is an expected shortfall measure, computed as the expected shortfall of the portfolio wealth minus the capital guarantees. Therefore we use the piecewise linear risk measure given in (9.18) but with only one linear function and a slope of 5. The expected shortfall is not only used at the terminal date of the optimization but at all time-steps in between. In our strategy we compute the expected shortfall for one quarter, three quarters and eight quarters in advance. 9.4.4

Results of the Historical Out-of-Sample Backtest

The out-of-sample test starts on 1 June 1996 and ends on 1 January 2005 with a quarterly frequency. The results are simulated results with all the inherent weaknesses and deviations from actual policies of large financial institutions. The statistics of the out-of-sample test for the portfolios and the assets are shown in Table 9.4. The asset and the portfolio evolutions throughout the historical out-of-sample test are shown in Figure 9.5. The graph shows that both portfolios would have a relatively steady evolution throughout the historical backtest with only one longer drawdown period between the third quarter in 2000 until the third quarter in 2002. The largest loss occurs in the third quarter of 1998, where the portfolio (9.1) loss is 16.3). The initial investments are mostly into the EU bond market and the subsequent portfolio gain allows the system to invest more into the stock market between 1996 and 1998. In the bull market phase, the portfolio with lower return guarantee would have had a higher return, since it could invest more into the risky assets. This higher allocation into the stock and bond market arises from the higher TABLE 9.4 Summary Statistics of the 4 Indices from 1 June 1996 to 1 January 2005 and Results of the Normality Tests for the Standard Residuals. A 0 Indicates that We Cannot Reject the Normality Assumption and a 1 Indicates the Rejection of Normality




Time series

r())

s())

SR

krt

skw

J /B (5))

LF (5))

Swiss stock market Swiss bond market EU stock market in CHF EU bond market in CHF EU stock market hedged EU bond market hedged 3-month LIBOR (SNB) Portfolio (1) (4) guarantee) Portfolio (2) (2.5) guarantee)

7.0 5.5 9.1 9.2 10.0 9.4 1.41 7.2 5.7

24.6 4.5 26.2 7.4 22.1 6.1  9.5 11.3

0.22 0.88 0.28 1.04 0.39 1.29  0.6 0.4

1.5 0.9 0.5 0.7 0.0 0.3  3.4 4.5

0.9 0.3 0.8 0.0 0.5 0.3  1.4 0.9

1 0 1 0 1 0 0 1 1

1 0 1 0 1 0 0 1 1

SOLVING ALM PROBLEMS VIA SEQUENTIAL STOCHASTIC PROGRAMMING

j

215

Portfolio (1) (4.0%) Guarantee (4.0%) Portfolio (2) (2.5%) Guarantee (2.5%) Swiss stock Swiss bond EU stock EU bond

260

240

Index Value

220

200

180

160

140

120

100

Apr97

Sep98

Jan00

May01

Oct02

Feb04

FIGURE 9.5 Results of the out-of-sample test for the Swiss case study and comparison to the asset classes.

distance to the minimum guarantee barrier and thus constrains the allocation less in this market phase. The evolution of the asset allocations of portfolio (1) are shown in Figure 9.6 as percentages of portfolio value. The large loss in the third quarter of 1998 leads to a dramatic increase of the money market investments and a sharp decrease of the investments into the Swiss stock market. The capacity to incur losses is reduced at this moment and the risky investments are consequently reduced. A similar behaviour can be seen during the drawdown from 2000 to early 2003, where the fund invests mostly into Swiss bonds and non-hedged EU bonds. Large changes in the asset allocation happen usually after significant changes in the portfolio value or after significant changes in the risk-return perception of the assets. In this phase (2000
2003) the portfolio (2) has a longer and stronger drawdown than portfolio (1), since portfolio (1) is more constrained by the distance to the minimal return guarantee. Also, the higher return guarantee forces the allocation to be more conservative and, therefore, limits the losses in this phase. When the stock markets are rising again (mid 2003
2005), portfolio (1) increases but not as strongly as portfolio (2). Therefore, we can conclude that the return guarantee acts as a security measure that limits losses in unfavourable times but has significant opportunity costs in rising markets. The dynamic changes in the asset allocation hold the wealth above the guarantee barrier, but would be impossible to be implemented in reality by a large /

/

216 j CHAPTER 9 Money Market Investments

Domestic Investments

(%) of Wealth

(%) of Wealth

Stock market

60 40 20 0

60

Bond market

40 20

Sep98

May01

0

Feb04

Sep98

Unhedged (CHF) Unhedged (CHF)

30

Hedged (CHF)

20 10

Sep98

May01

Feb04

Hedged (CHF)

60 (%) of Wealth

(%) of Wealth

Feb04

EU Bond Market Investments

EU Stock Market Investments 40

0

May01

40 20 0

Sep98

May01

Feb04

FIGURE 9.6 Asset allocation of portfolio (1) in percentage of wealth.

financial institution. This behaviour is induced by the type of risk measure and the factor model. Institutions stick much closer to their strategic allocation and change their allocation less dramatically. This method could, however, provide a guide for tactical asset allocation. Most of the stock investments are hedged as shown in Figure 9.6. The risk-return tradeoff for stocks seems to be more acceptable with hedging than without hedging. Moreover, most of the investment in stocks occurs before 2001 and in this period the Swiss franc is constantly gaining in value and thus reducing returns from the international investments. The Swiss franc loses value after mid-2002 when most of the international bond investment occurs. Therefore, most of the bond investment takes place without hedging. By introducing the international assets twice as either hedged or not hedged, we use the standard portfolio optimization to make the hedging decision in parallel with the portfolio construction. However, a historical backtest starting after 2000 would have much more difficulties to remain above the barriers, since the two least risky assets (Swiss money market and bond market) did not yield returns above 4). For this reason, many Swiss pension funds came into a situation of severe financial stress. This was one of the reasons why the capital guarantee has been strongly reduced and is adjusted biennially according to market expectations. Despite the relatively high transaction costs, the performance of portfolio (1) is satisfactory with an average return of 7.1). The performance is similar to the Swiss stock market but markedly higher than that of the Swiss bond market or money market accounts.

SOLVING ALM PROBLEMS VIA SEQUENTIAL STOCHASTIC PROGRAMMING

j

217

The standard deviation is moderate, but much smaller than any of the stock market investment possibilities. Hence, the Sharpe ratio lies between the one for the stock and the bond markets. The performance and the risk numbers for portfolio (2) are not as good, but a backtest that would have started earlier, e.g. 1990, would have shown better results. Both portfolios show the strengths and weaknesses of performance guarantees for funds. Often guarantees are only judged with respect to the opportunity costs they carry. In this case study, since it includes a strong bull and bear market, the positive sides are also highlighted. However, the very dynamic changes in the allocation would be very difficult to follow by a large financial institution which limits this backtest to be only a demonstration of the proposed method.

9.5

CONCLUSION AND OUTLOOK

The first part of the chapter shows that we can approximately solve dynamic programming problems for dynamic portfolio management problems. We prove that by approximating the true dynamics by a set of scenarios and re-solving the problem at every time-step, we solve the dynamic programming problem with an arbitrarily objective function error. Future work should address the question of computational efficiency versus other established method. In the second part of the chapter, we describe asset return and portfolio dynamics. We argue that the most suitable risk measures for ALM situations or guaranteed return funds are shortfall risk measures. A penalty function classifies the shortfall of the assets with respect to the liabilities where large ‘non-achievements’ are more severely penalized than small ‘non-achievements’ on future scenarios where the minimal return is not achieved. The optimization problem is solved with the aim of maximizing the return above the guarantee over the planning horizon, while keeping the shortfall risk below a predetermined limit. In this case study, the risk aversion varies throughout the historical backtest and depends on the distance to the barrier. The optimization always reduces the risk exposure when the portfolio wealth moves closer to the barrier and increases the risk exposure when the portfolio moves away from the barrier. In this way, we introduce a feedback from the portfolio results to the current portfolio decisions and adapt the risk aversion to the loss incurring capacity. The backtesting results will be compared in future research to other well known ALM techniques such as immunization or fixed-mix allocations. These comparisons could further illustrate the strength and weakness of this approach.

ACKNOWLEDGEMENTS The authors wish to acknowledge the comments of three anonymous referees and thank the Editor-in-Chief for his detailed comments and corrections which materially improved the chapter.

218 j CHAPTER 9

REFERENCES Alexander, C., Market Models, 2001 (Wiley: Chichester). Artzner, P., Delbaen, F., Eber, J. and Heath, D., Coherent measures of risk. Math. Finance, 1999, 9(3), 203
/228. Batina, I., Stoorvogel, A.A. and Weiland, S., Model predictive control of linear stochastic systems with state and input constraints. In Conference on Decision and Control pp. 1564
/1569, 2002 (Las Vegas, NV). Bertsekas, D.P., Dynamic Programming and Optimal Control, Vol. I, 1995 (Althena Scientific: Belmont, MA). Bertsekas, D.P. and Shreve, S.E., Stochastic Optimal Control: The Discrete-Time Case, 1978 (Academic Press: New York). Birge, J.R., Donohue, C., Holmes, D. and Svintsitski, O., A parallel implementation of the nested decomposition algorithm for multistage stochastic linear programs. Math. Program., 1994, 75, 327
/352. Bratley, P. and Fox, B., Algorithm 659: Implementing Sobol’s quasirandom sequence generator ACM Trans. Math. Software, 1988, 14(1), 88
/100. Brennan, M.J., Schwartz, E.S. and Lagnado, R., Strategic asset allocation J. Econ. Dyn. Control, 1997, 21(8
/9), 1377
/1403. Carino, D.R. and Ziemba, W.T., Formulation of the Russell
/Yasuda Kasai financial planning model Operat. Res., 1998, 46(4), 433
/449. Casella, G. and Berger, R.L., Statistical Inference, 2nd edn, 2002 (Duxbury: Pacific Grove, CA). Dempster, M.A.H., Sequential importance sampling algorithms for dynamic stochastic programming. Trans. St. Petersburg Steklov Math. Inst., 2004, 312, 94
/129. Dempster, M.A.H. and Consigli, G., The CALM stochastic programming model for dynamic asset and liability management. In Worldwide Asset and Liability Modeling, edited by W. Ziemba and J. Mulvey, pp. 464
/500, 1998 (Cambridge University Press: Cambridge). Dempster, M.A.H., Germano, M., Medova, E., Rietbergen, M., Sandrini, F. and Scrowston, M., Managing guarantees. J. Portfolio Manag., 2006, 32(2), 51
/61. Dempster, M.A.H., Germano, M., Medova, E., Rietbergen, M.I., Sandrini, F. and Scrowston, M., Designing minimum guaranteed return funds. Quant. Finance, 2007, 7(2), (this issue). Dempster, M.A.H. and Thompson, R., EVPI-based importance sampling solution procedures for multistage stochastic linear programmes on parallel MIMD architectures. Ann. Operat. Res., 1999, 90, 161
/184. Dondi, G., Models and dynamic optimisation for the asset and liability management for pension funds. PhD thesis, Dissertation ETH No. 16257, ETH Zurich, 2005. Dondi, G., Herzog, F., Schumann, L.M. and Geering, H.P., Dynamic asset & liability management for Swiss pension funds. In Handbook of Asset and Liability Management, Vol. 2, edited by S. Zenios and W. Ziemba, 2007 (North-Holland: Oxford) in press. Dupacova, J., Consigli, G. and Wallace, S., Scenario for multistage stochastic programs. Ann. Operat. Res., 2000, 100, 25
/53. Fabozzi, F.J., (Ed.), The Handbook of Fixed Income Securities, 7th edn, 2005 (McGraw-Hill: New York). Geyer, A., Herold, W., Kontriner, K. and Ziemba, W.T., The Innovest Austrian pension fund financial planning model: InnoALM. In Conference on Asset and Liability Management: From Institutions to Households, May 2004 (Nicosia). Givens, G. and Hoeting, J., Computational Statistics, 2005 (Wiley Interscience: Hoboken). Glasserman, P., Monte Carlo Methods in Financial Engineering, 2004 (Springer: New York). Hoyland, K., Kaut, M. and Wallace, S., A heuristic for moment-matching scenario generation. Comput. Optim. Appl., 2003, 24, 169
/185.

SOLVING ALM PROBLEMS VIA SEQUENTIAL STOCHASTIC PROGRAMMING

j

219

Illien, M., Optimal strategies for currency portfolios. Diploma thesis, IMRT. Introduction to Stochastic Programming, 1997 (Springer: New York). McNeil, A.J., Frey, R. and Embrechts, P., Quantitative Risk Management: Concepts, Techniques and Tools, 2005 (Princeton University Press: Princeton, NJ). Mulvey, J., Gould, G. and Morgan, C., An asset and liability management system for Towers PerrinTillinghast. Interfaces, 2000, 30, 96
/114. Mulvey, J.M., Generating scenarios for the Towers Perrin investment system. Interfaces, 1995, 26, 1
/15. Mulvey, J.M. and Shetty, B., Financial planning via multi-stage stochastic optimization. Comput. Operat. Res., 2004, 31, 1
/20. Pennanen, T., Epi-convergent discretizations of multistage stochastic programs. Math. Operat. Res., 2005, 30, 245
/256. Pennanen, T. and Koivu, M., Epi-convergent discretizations of stochastic programs via integration quadratures. Numer. Math., 2005, 100, 141
/163. Riedel, F., Dynamic coherent risk measures. Stochast. Process. Appl., 2004, 112, 185
/200. Rockafellar, R. and Uryasev, S., Optimization of conditional value-at-risk. J. Risk, 2000, 2(3), 21
/41. Rockafellar, R. and Uryasev, S.P., Conditional value-at-risk for general loss distributions. J. Bank. Finance, 2002, 26, 1443
/1471. Ro¨misch, W., Stability of stochastic programming problems. In Handbooks in Operations Research and Management Science, Vol. 10, edited by A. Ruszczynski and A. Shapiro, pp. 483
/554, 2003 (Elsevier: Amsterdam). Zenios, S.A. and Ziemba, W.T., (Eds), Handbook of Asset and Liability Management, Vol. 1, 2007 (Oxford: North-Holland), in press. Ziemba, W.T., The Stochastic Programming Approach to Asset Liability and Wealth Management, 2003 (AIMR: Charlottesville, VA). Ziemba, W.T. and Rockafellar, T., Modified risk measures and acceptance sets. In Working Paper, 2000 (Sauder School of Business, University of British Columbia).

APPENDIX 9.A: ADDITIONAL DATA FOR THE CASE STUDY In Table 9.A1 we give the factors of the case study reported. Factors that were never selected, such as the dividend yield or macroeconomic variables (GNP), are omitted from the table. Table 9.A1 All Factors for Case Study 3 with Swiss and EU Data Factor no.

Factor name

1 2 3 4 5 6 7 8 9 10 11 12

10-year Swiss Government bond interest rate 10-year EU Benchmark Government bond (GB) interest rate log (EP ratio)  log (10-year GB int. rate) Switzerland log (EP ratio)  log (10-year GB int. rate) EU 10-year GB rate  3-month LIBOR Switzerland 10-year GB rate  3-month LIBOR EU FX sopt rate CHFuEuro (DM) 3 months momentum. Swiss stock market 3 months momentum. EU stock market 3 months momentum. Swiss bond market 3 months momentum. EU bond market FX 3-monthforward rate CHFuEuro (DM)

220 j CHAPTER 9

APPENDIX 9.B: DYNAMIC PROGRAMMING RECURSION FOR THE SAMPLE APPROXIMATION Proof of Theorem 9.1: Let Ps ðtÞ :¼ ½ps ðtÞ; ps ðt þ 1Þ; . . . ; ps ðT  1Þ and insert (9.3) into (9.5) which yields ( " #) T 1 X s s s E^ Lði; y ðiÞ; p ðiÞÞ þ MðT ; p ðT ÞÞ J ðt; yðtÞÞ ¼ max s ^s

P ðtÞ

¼

i¼t

max

ps ðtÞ;Ps ðtþ1Þ

   T 1 X s s s s s E^ Lðt; y ðtÞ; p ðtÞÞ þ Lði; y ðiÞ; p ðiÞÞ þ MðT ; p ðT ÞÞ i¼tþ1

s.t. y s ðt þ 1Þ ¼ Dðt; y s ; ps Þ þ Sðt; y s ; ps ÞEs ðtÞ; ð9:B1Þ and nonanticipativity. Since L(t, ys(t), ps(t)) is independent of the future decisions Ps(t  1) and the scenarios for es(i), i  t are independent of ps(t), move the maximization operator over Ps(t  1) inside the bracket to obtain ( " J^s ðt; yðtÞÞ ¼ max E^ Lðt; y s ðtÞ; ps ðtÞÞ ps ðtÞ

( " #)#) T 1 X s s s þ max E^ Lði; y ðiÞ; p ðiÞÞ þ MðT ; p ðT ÞÞ s P ðtþ1Þ

i¼tþ1

n h io ¼ max E^ Lðt; y s ðtÞ; ps ðtÞÞ þ J^s ðt þ 1; y s ðt þ 1ÞÞ ; ps ðtÞ

subject to the dynamical constraints and nonanticipativity. From the first of the dynamical constraints ys(t  1)  D(t, ys, ps)  S(t, ys, ps)es(t) and the basic idea of Bellman’s principle ( " #) T 1 X E^ Lði; y s ðiÞ; ps ðiÞÞ þ MðT ; ps ðT ÞÞ ; J^s ðt þ 1; yðt þ 1ÞÞ ¼ max s P ðtþ1Þ

i¼tþ1

subject to the remaining dynamical constraints and nonanticipativity, to give n h io ^ Lðt; y s ðtÞ; ps ðtÞÞ þ J^s ðt þ 1; Dðt; y s ; ps Þ þ Sðt; y s ; ps Þ Es ðtÞÞ E J^s ðt; yðtÞÞ ¼ max ps ðtÞ n h io ^ Lðt; y s ðtÞ; us ðtÞÞ þ J^s ðt þ 1; Dðt; y s ; us Þ þ Sðt; y s ; us Þ Es ðtÞÞ ; E ¼ max s u ðtÞ2U

subject to the appropriate dynamical constraints and nonanticipativity, where we convert the maximization over ps(t) to a maximization over us(t), using the fact that for any

SOLVING ALM PROBLEMS VIA SEQUENTIAL STOCHASTIC PROGRAMMING

j

221

function f of x and u it is true that maxf f ðx; pðxÞÞg ¼ maxf f ðx; uÞg; p2Q

u2U

where Q is the set of all functions p(x) such that pðxÞ 2 U 8x. This statement can be found in Bertsekas (1995, Chapter 2). I

CHAPTER

10

Designing Minimum Guaranteed Return Funds M. A. H. DEMPSTER, M. GERMANO, E. A. MEDOVA, M. I. RIETBERGEN, F. SANDRINI and M. SCROWSTON CONTENTS 10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2 Stochastic Optimization Framework . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2.1 Set Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2.2 Model Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2.3 Objective Functions: Expected Average Shortfall and Expected Maximum Shortfall . . . . . . . . . . . . . . . . . . . . . . . 10.3 Bond Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3.1 Yield Curve Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3.2 Pricing Coupon-Bearing Bonds . . . . . . . . . . . . . . . . . . . . . . . . 10.4 Historical Backtests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.5 Robustness of Backtest Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10.1

I

223 225 225 226 229 232 232 234 235 240 242 242 243

INTRODUCTION

of investment products aimed at attracting investors who are worried about the downside potential of the financial markets for pension investments. The main feature of these products is a minimum guaranteed return together with exposure to the upside movements of the market. There are several different guarantees available in the market. The most common one is the nominal guarantee which guarantees a fixed percentage of the initial investment. However there also exist funds with a guarantee in real terms which is linked to an inflation index. Another distinction can be made between fixed and flexible guarantees, with the N RECENT YEARS THERE HAS BEEN A SIGNIFICANT GROWTH

223

224 j CHAPTER 10

fixed guarantee linked to a particular rate and the flexible to, for instance, a capital market index. Real guarantees are a special case of flexible guarantees. Sometimes the guarantee of a minimum rate of return is even set relative to the performance of other pension funds. Return guarantees typically involve hedging or insuring. Hedging involves eliminating the risk by sacrificing some or all of the potential for gain, whereas insuring involves paying an insurance premium to eliminate the risk of losing a large amount. Many government and private pension schemes consist of defined benefit plans. The task of the pension fund is to guarantee benefit payments to retiring clients by investing part of their current wealth in the financial markets. The responsibility of the pension fund is to hedge the client’s risk, while meeting the solvency requirements in such a way that all benefit payments are met. However at present there are significant gaps between fund values, contributions made by employees and pension obligations to retirees. One way in which a guarantee can be achieved is by investing in zero-coupon Treasury bonds with a maturity equal to the time horizon of the investment product in question. However using this option foregoes all upside potential. Even though the aim is protect the investor from the downside, a reasonable expectation of returns higher than guaranteed needs to remain. In this chapter we will consider long-term nominal minimum guaranteed return plans with a fixed time horizon. They will be closed end guarantee funds; after the initial contribution there is no possibility of making any contributions during the lifetime of the product. The main focus will be on how to optimally hedge the risks involved in order to avoid having to buy costly insurance. However, this task is not straightforward, as it requires long-term forecasting for all investment classes and dealing with a stochastic liability. Dynamic stochastic programming is the technique of choice to solve this kind of problem as such a model will automatically hedge current portfolio allocations against the future uncertainties in asset returns and liabilities over a long horizon (see e.g. Consigli and Dempster 1998; Dempster et al. 2000, 2003). This will lead to more robust decisions and previews of possible future benefits and problems contrary to, for instance, static portfolio optimization models such as the Markowitz (1959) mean-variance allocation model. Consiglio et al. (2007) have studied fund guarantees over single investment periods and Hertzog et al. (2007) treat dynamic problems with a deterministic risk barrier. However, a practical method should have the flexibility to take into account multiple time periods, portfolio constraints such as prohibition of short selling and varying degrees of risk aversion. In addition, it should be based on a realistic representation of the dynamics of the relevant factors such as asset prices or returns and should model the changing market dynamics of risk management. All these factors have been carefully addressed here and are explained further in the sequel. The rest of the chapter is organized as follows. In Section 10.2 we describe the stochastic optimization framework, which includes the problem set up, model constraints and possible objective functions. Section 10.3 presents a three-factor term structure model and its application to pricing the bond portfolio and the liability side of the fund on individual scenarios. As our portfolio will mainly consist of bonds, this area has been extensively

DESIGNING MINIMUM GUARANTEED RETURN FUNDS

j

225

researched. Section 10.4 presents several historical backtests to show how the framework would have performed had it been implemented in practice, paying particular attention to the effects of using different objective functions and varying tree structures. Section 10.5 repeats the backtest when the stock index is modelled as a jumping diffusion so that the corresponding returns exhibit fat tails and Section 10.6 concludes. Throughout this chapter boldface is used to denote random entities.

10.2

STOCHASTIC OPTIMIZATION FRAMEWORK

In this section we describe the framework for optimizing minimum guaranteed return funds using stochastic optimization. We will focus on risk management as well as strategic asset allocation concerned with allocation across broad asset classes, although we will allow specific maturity bond allocations. 10.2.1

Set Up

This section looks at several methods to optimally allocate assets for a minimum guaranteed return fund using expected average and expected maximum shortfall risk measures relative to the current value of the guarantee. The models will be applied to eight different assets: coupon bonds with maturity equal to 1, 2, 3, 4, 5, 10 and 30 years and an equity index, and the home currency is the euro. Extensions incorporated into these models are the presence of coupon rates directly dependent on the term structure of bond returns and the annual rolling over of the coupon-bearing bonds. We consider a discrete time and space setting. The time interval considered is given by f0; ð1=12Þ; ð2=12Þ; . . . ; T g, where the times indexed by t
0,1, . . . , T1 correspond to decision times at which the fund will trade and T to the planning horizon at which no decision is made (see Figure 10.1). We will be looking at a five-year horizon. Uncertainty V is represented by a scenario tree, in which each path through the tree corresponds to a scenario v in V and each node in the tree corresponds to a time along one or more scenarios. An example scenario tree is given in Figure 10.2. The probability p(v) of scenario v in V is the reciprocal of the total number of scenarios as the scenarios are generated by Monte Carlo simulation and are hence equiprobable. The stock price process S is assumed to follow a geometric Brownian motion, i.e. dSt St

¼ mS dt þ sS dWSt ;

ð10:1Þ

Stages: s=1

t=0

s=2

t=1/12 t=2/12 t=3/12 t=4/12

t=5/12

Time

FIGURE 10.1 Time and stage setting.

t=1/2

t=7/12 t=8/12 t=9/12 t=10/12 t=11/12

t=1

t=13/12 t=14/12 t=15/12

226 j CHAPTER 10

t=0

t=1/4

t=1/2

s=1

t=3/4

t=1

t=5/4

t=3/2

t=7/4

s=2

t=2 s=3

FIGURE 10.2 Graphical representation of scenarios.

where dWSt is correlated with the dWt terms driving the three term structure factors discussed in Section 10.3. 10.2.2

Model Constraints

Let (see Table 10.1) “

ht(v) denote the shortfall at time t and scenario v, i.e. ht ðoÞ :¼ maxð0; Lt ðoÞ  Wt ðoÞÞ

“

“

8o 2 O t 2 T total :

ð10:2Þ

H ðoÞ :¼ maxt2T total ht ðoÞ denote the maximum shortfall over time for scenario v. The constraints considered for the minimum guaranteed return problem are: Cash balance constraints. These constraints ensure that the net cash flow at each time and at each scenario is equal to zero X

buy

þ fP0;a ðoÞx0;a ðoÞ ¼ W0 ;

o 2 O;

ð10:3Þ

a2A

X 1 a X  sell  dt1 ðoÞF a xt;a ðoÞ þ gPt;a ðoÞxt;a ðoÞ a2A a2AnfS g 2 X buy þ fPt;a ðoÞxt;a ðoÞ; o 2 O t 2 T d nf0g: ¼

ð10:4Þ

a2A

In (10.4) the left-hand side represents the cash freed up to be reinvested at time t 2 T d nf0g and consists of two distinct components. The first term represents the semi-annual coupons received on the coupon-bearing Treasury bonds held between time t1 and t, the second term represents the cash obtained from selling part of the portfolio. This must equal the value of the new assets bought given by the right-hand side of (10.4).

DESIGNING MINIMUM GUARANTEED RETURN FUNDS

j

227

TABLE 10.1 Variables and Parameters of the Model Time sets   T total ¼ 0; 121 ; . . . ; T d ¼ f0; 1; . . . ; T  1g total d Ti¼T  nT  T c ¼ 12 ; 32 ; . . . ; T  12

Set of all times considered in the stochastic program Set of decision times Set of intermediate times Times when a coupon is paid out in-between decision times

Instruments St(v) BtT ðoÞ T dBt ðoÞ

Dow Jones Eurostoxx 50 index level at time t in scenario v EU Treasury bond with maturity T at time t in scenario v coupon rate of EU Treasury bond with maturity T at time t in scenario v face value of EU Treasury bond with maturity T EU zero-coupon Treasury bond price at time t in scenario v

T

FB Zt(v) Risk management barrier yt,T(v)

EU zero-coupon Treasury yield with maturity T at time t in scenario v Annual guaranteed return Nominal barrier at time t in scenario v

G LtN ðoÞ Portfolio evolution A buy sell ðoÞ=Pt;a ðoÞ Pt;a fg xt,a(v)

Set of all assets Buy/sell price of asset a
A at time t in scenario v Transaction costs for buying/selling Quantity held of asset a
A between time t and t  1/12 in scenario v Quantity bought/sold of asset a
A at time t in scenario v Initial portfolio wealth Portfolio wealth before rebalancing at time t
T in scenario v Portfolio wealth after rebalancing at time t
T c [ T d nfT g in scenario v Shortfall at time t in scenario v

þ  xt;a ðoÞ=xt;a ðoÞ W0 Wt (v) wt (v)

ht ðoÞ :¼ maxð0; Lt ðoÞ  Wt ðoÞÞ

“

Short sale constraints. In our model we assume no short selling of any stocks or bonds xt;a ðoÞ  0; þ ðoÞ  0; xt;a

 ðoÞ  0; xt;a

“

a 2 A;

o 2 O;

t 2 T total ;

ð10:5Þ

8a 2 A;

8o 2 O;

8t 2 T total nfT g;

ð10:6Þ

8a 2 A;

8o 2 O;

8t 2 T total nf0g:

ð10:7Þ

Information constraints. These constraints ensure that the portfolio allocation cannot be changed during the period from one decision time to the next and hence that no decisions with perfect foresight can be made þ  ðoÞ ¼ xt;a ðoÞ ¼ 0; xt;a

a 2 A; o 2 O; t 2 T i nT c :

ð10:8Þ

228 j CHAPTER 10 “

Wealth constraint. These constraints determine the portfolio wealth at each point in time prior to and after rebalancing wt ðoÞ ¼

X

buy Pt;a ðoÞxt;a ðoÞ;

o2O

t 2 T total nfT g;

ð10:9Þ

a2A

Wt ðoÞ ¼

X

sell Pt;a ðoÞxtð1=12Þ;a ðoÞ;

o 2 O t 2 T total nf0g;

ð10:10Þ

a2A

wT ðoÞ ¼

X

gPTsell;a ðoÞxT ð1=12Þ;a ðoÞ þ

a2A

X 1 a dT 1 ðoÞF a xT ð1=12Þ;a ðoÞ; a2AnfS g 2

o 2 O: ð10:11Þ

“

Accounting balance constraints. These constraints give the quantity invested in each asset at each time and for each scenario þ ðoÞ; x0;a ðoÞ ¼ x0;a

þ  ðoÞ  xt;a ðoÞ; xt;a ðoÞ ¼ xt1=12;a ðoÞ þ xt;a

a 2 A; o 2 O;

ð10:12Þ

a 2 A; o 2 O; t 2 T total nf0g: ð10:13Þ

“

The total quantity invested in asset a
A between time t and t  (1u12) is equal to the total quantity invested in asset a
A between time t(1u12) and t plus the quantity of asset a
A bought at time t minus the quantity of asset a
A sold at time t. Annual rolling constraint. This constraint ensures that at each decision time all the coupon-bearing Treasury bond holdings are sold  ðoÞ ¼ xtð1=12Þ;a ðoÞ; xt;a

“

a 2 AnfSg; o 2 O; t 2 T d nf0g:

ð10:14Þ

Coupon re-investment constraints. We assume that the coupon paid every six months will be re-invested in the same coupon-bearing Treasury bond þ ðoÞ ¼ xt;a

ð1=2Þdat ðoÞF a xtð1=12Þ;a ðoÞ buy fPt;a ðoÞ

;

 ðoÞ ¼ 0; xt;a

þ  xt;S ðoÞ ¼ xt;S ðoÞ ¼ 0;

a 2 AnfS g; o 2 O; t 2 T c :

ð10:15Þ

DESIGNING MINIMUM GUARANTEED RETURN FUNDS “

j

229

Barrier constraints. These constraints determine the shortfall of the portfolio at each time and scenario as defined in Table 10.1 ht ðoÞ þ Wt ðoÞ  Lt ðoÞ;

ht ðoÞ  0;

o 2 O; t 2 T total ;

o 2 O; t 2 T total :

ð10:16Þ

ð10:17Þ

As the objective of the stochastic program will put a penalty on any shortfall, optimizing will ensure that ht (v) will be zero if possible and as small as possible otherwise, i.e. ht ðoÞ ¼ maxð0; Lt ðoÞ  Wt ðoÞÞ;

o 2 O; t 2 T total ;

ð10:18Þ

which is exactly how we defined ht (v) in (10.2). To obtain the maximum shortfall for each scenario, we need to add one of the following two sets of constraints: HðoÞ  ht ðoÞ;

H ðoÞ  ht ðoÞ;

o 2 O; t 2 T d [ fT g

ð10:19Þ

o 2 O; t 2 T total :

ð10:20Þ

Constraint (10.19) needs to be added if the max shortfall is to be taken into account on a yearly basis and constraint (10.20) if max shortfall is calculated on a monthly basis. 10.2.3 Objective Functions: Expected Average Shortfall and Expected Maximum Shortfall Starting with an initial wealth W0 and an annual nominal guarantee of G, the liability at the planning horizon at time T is given by W0 ð1 þ G ÞT :

ð10:21Þ

To price the liability at time t B T consider a zero-coupon Treasury bond which pays 1 at time T, i.e. ZT (v)
1, for all scenarios v
V. The zero-coupon Treasury bond price at time t in scenario v assuming continuous compounding is given by Zt ðoÞ ¼ eyt;T ðoÞðT t Þ ;

ð10:22Þ

where yt,T (v) is the zero-coupon Treasury yield with maturity T at time t in scenario v. This gives a formula for the value of the nominal (fixed) guarantee barrier at time t in scenario v as

230 j CHAPTER 10

LtN ðoÞ :¼ W0 ð1 þ GÞT Zt ðoÞ ¼ W0 ð1 þ GÞT eyt;T ðoÞðT tÞ :

ð10:23Þ

In a minimum guaranteed return fund the objective of the fund manager is twofold; firstly to manage the investment strategies of the fund and secondly to take into account the guarantees given to all investors. Investment strategies must ensure that the guarantee for all participants of the fund is met with a high probability. In practice the guarantor (the parent bank of the fund manager) will ensure the investor guarantee is met by forcing the purchase of the zero coupon bond of (10.22) when the fund is sufficiently near the barrier defined by (10.23). Since all upside potential to investors is thus foregone, the aim of the fund manager is to fall below the barrier with acceptably small if not zero probability. Ideally we would add a constraint limiting the probability of falling below the barrier in a VaR-type minimum guarantee constraint, i.e.   h ð o Þ > 0

a P max t total

ð10:24Þ

t2T

for a small. However, such scenario-based probabilistic constraints are extremely difficult to implement, as they may without further assumptions convert the convex large-scale optimization problem into a non-convex one. We therefore use the following two convex approximations in which we trade off the risk of falling below the barrier against the return in the form of the expected sum of wealth. Firstly, we look at the expected average shortfall (EAS) model in which the objective function is given by (

 ) ht ðoÞ pðoÞ ð1  bÞWt ðoÞ  b n max o jT d [ T j xt;a ðoÞ;x þ ðoÞ;x  ðoÞ: o2O t2T d [T t;a t ;a a2A;o2O;t2T d [T

X X (

!) ht ðoÞ : pðoÞ Wt ðoÞ  b pðoÞ ¼ n max o ð1  bÞ d xt;a ðoÞ;x þ ðoÞ;x  ðoÞ: o2O o2O t;a t;a t2T d [T t2T d [T jT [ T j X

X

!

X

X

a2A;o2O;t2T d [T

ð10:25Þ In this case we maximize the expected sum of wealth over time while penalizing each time the wealth falls below the barrier. For each scenario v
V we can calculate the average shortfall over time and then take expectations over all scenarios. In this case only shortfalls at decision times are taken into account and any serious loss in portfolio wealth in-between decision times is ignored. However, from the fund manager’s and guarantor’s perspective the position of the portfolio wealth relative to the fund’s barrier is significant on a continuous basis and serious or repeated drops below this barrier might force the purchase of expensive insurance. To capture this feature specific to minimum guaranteed return funds, we also consider an objective function in which the shortfall of the portfolio is considered on a monthly basis.

j

DESIGNING MINIMUM GUARANTEED RETURN FUNDS

231

For the expected average shortfall with monthly checking (EAS MC) model the objective function is given by (

X

!

X

X

X h ðoÞ t ð1bÞ pðoÞ W ðoÞ b pðoÞ max n o t total j jT xt;a ðoÞ;x þ ðoÞ;x  ðoÞ: o2O o2O t;a t;a t2T d [T t2T total

!) : ð10:26Þ

a2A;o2O;t2T d [T

Note that although we still only rebalance once a year shortfall is now being measured on a monthly basis in the objective and hence the annual decisions must also take into account the possible effects they will have on the monthly shortfall. The value of 0 5b51 can be chosen freely and sets the level of risk aversion. The higher the value of b, the higher the importance given to shortfall and the less to the expected sum of wealth, and hence the more risk-averse the optimal portfolio allocation will be. The two extreme cases are represented by b
0, corresponding to the ‘unconstrained’ situation, which is indifferent to the probability of falling below the barrier, and b
1, corresponding to the situation in which the shortfall is penalized and the expected sum of wealth ignored. In general short horizon funds are likely to attract more risk-averse participants than long horizon funds, whose participants can afford to tolerate more risk in the short run. This natural division between short and long-horizon funds is automatically incorporated in the problem set up, as the barrier will initially be lower for long-term funds than for short-term funds as exhibited in Figure 10.3. However, the importance of closeness to the barrier can be adjusted by the choice of b in the objective. The second model we consider is the expected maximum shortfall (EMS) model given by ( n

max

þ  xt;a ðoÞ;xt;a ðoÞ;xt;a ðoÞ: d a2A;o2O;t2T [fT g

o ð1  bÞ

X o2O

pðoÞ

!

X

Wt ðoÞ  b

t2T d [fT g

X

!) pðoÞHðoÞ

o2O

using the constraints (10.19) to define H(v). 115 110 105 100 95 1 Year barrier 5 Year barrier

90 85 80 1-Jan-99

1-Jan-00

31-Dec-00

31-Dec-01

31-Dec-023

FIGURE 10.3 Barrier for one-year and five-year 2) guaranteed return fund.

1-Dec-03

ð10:27Þ

232 j CHAPTER 10

For the expected maximum shortfall with monthly checking (EMS MC) model the objective function remains the same but H(v) is now defined by (10.20). In both variants of this model we penalize the expected maximum shortfall, which ensures that H(v) is as small as possible for each scenario v
V. Combining this with constraints (10.19)/(10.20) it follows that H(v) is exactly equal to the maximum shortfall. The constraints given in Section 10.2.2 apply to both the expected average shortfall and expected maximum shortfall models. The EAS model incurs a penalty every time portfolio wealth falls below the barrier, but it does not differentiate between a substantial shortfall at one point in time and a series of small shortfalls over time. The EMS model on the other hand, focuses on limiting the maximum shortfall and therefore does not penalize portfolio wealth falling just slightly below the barrier several times. So one model limits the number of times portfolio wealth falls below the barrier while the other limits any substantial shortfall.

10.3

BOND PRICING

In this section we present a three-factor term structure model which we will use to price both our bond portfolio and the fund’s liability. Many interest-rate models, like the classic one-factor Vasicek (1977) and Cox et al. (1985) class of models and even more recent multi-factor models like Anderson and Lund (1997), concentrate on modelling just the short-term rate. However, for minimum guaranteed return funds we have to deal with a long-term liability and bonds of varying maturities. We therefore must capture the dynamics of the whole term structure. This has been achieved by using the economic factor model described below in Section 10.3.1. In Section 10.3.2 we describe the pricing of coupon-bearing bonds and Section 10.3.3 investigates the consequences of rolling the bonds on an annual basis. 10.3.1

Yield Curve Model

To capture the dynamics of the whole term structure, we will use a Gaussian economic factor model (EFM) (see Campbell 2000 and also Nelson and Siegel 1987) whose evolution under the risk-neutral measure Q is determined by the stochastic differential equations dXt ¼ ðmX  lX Xt Þdt þ sX dWXt ;

ð10:28Þ

sY dWYt ;

ð10:29Þ

dYt ¼ ðmY  lY Yt Þdt þ

dRt ¼ k ðXt þ Yt  Rt Þdt þ

sR dWRt ;

ð10:30Þ

where the dW terms are correlated. The three unobservable Gaussian factors R, X and Y represent respectively a short rate, a long rate and the slope between an instantaneous short rate and the long rate. Solving these equations the following formula for the yield at time t with time to maturity equal to T t is obtained (for a derivation, see Medova et al. 2005) yt;T ¼

Aðt; T ÞRt þ Bðt; T ÞXt þ Cðt; T ÞYt þ Dðt; T Þ ; T

ð10:31Þ

DESIGNING MINIMUM GUARANTEED RETURN FUNDS

j

233

where Aðt; T Þ :¼

k Bðt; T Þ :¼ k  lX

k C ðt; T Þ :¼ k  lY  D ðt; T Þ :¼

1 1  ekðT t Þ ; k

ð10:32Þ

(

)

1

1 lX ðT t Þ k ðT t Þ  1e ; 1e lX k )



1 1 1  elY ðT t Þ  1  ek ðT t Þ ; lY k

ð10:33Þ

(

1 T  t  1  ek ðT t Þ k



mX lX

þ

mY

! 

lY

mX lX

B ðt; T Þ 

mY lY

ð10:34Þ

C ðt; T Þ

( 3 mX2 i

mY2

n2

1X 1  e2lX ðT t Þ þ i 1  e2lY ðT t Þ þ i 1  e2k ðT t Þ þ pi2 ðT  t Þ  2 i¼1 2lX 2lY 2k

2mXi ni

2mXi pi

2mXi mYi 1  eðlX þkÞðT t Þ þ 1  eðlX þlY ÞðT t Þ þ 1  elX ðT t Þ lX þ lY lX þ k lX 

2mYi pi

2ni pi

2mYi ni ðlY þk ÞðT t Þ lY ðT t Þ k ðT t Þ þ þ ð10:35Þ 1e 1e þ 1e lY þ k lY k þ

and mXi :¼  mYi :¼ 

ksXi lX ðk  lX Þ

;

ksYi

; lY ðk  lY Þ s Xi sYi sR ni :¼ þ  i; k  lX k  lY k   pi :¼  mXi þ mYi þ ni :

ð10:36Þ

Bond pricing must be effected under the risk-neutral measure Q. However, for the model to be used for forward simulation the set of stochastic differential equations must be adjusted to capture the model dynamics under the real-world or market measure P. We therefore have to model the market prices of risk which take us from the risk-neutral measure Q to the real-world measure P.

234 j CHAPTER 10

Under the market measure P we adjust the drift term by adding the risk premium given by the market price of risk g in terms of the quantity of risk. The effect of this is a change in the long-term mean, e.g. for the factor X the long-term mean now equals ðmX þ gX sX Þ=lX . It is generally assumed in a Gaussian world that the quantity of risk is given by the volatility of each factor. This gives us the following set of processes under the market measure dXt ¼ ðmX  lX Xt þ gX sX Þdt þ sX dWXt ; dYt ¼ ðmY  lY Yt þ gY sY Þdt þ sY dWYt ; dRt ¼ fkðXt þ Yt  Rt Þ þ gR sR gdt þ sR dWRt ;

ð10:37Þ ð10:38Þ ð10:39Þ

where all three factors contain a market price of risk g in volatility units. The yields derived in the economic factor model are continuously compounded while most yield data are annually compounded. So for appropriate comparison when estimating the parameters of the model we will have to convert the annually compounded yields into continuously compounded yields using the transformation y continuous ¼ lnð1 þ y annual Þ:

ð10:40Þ

In the limit as T tends to infinity it can be shown that expression (10.31) derived for the yield does not tend to the ‘long rate’ factor X, but to a constant. This suggests that the three factors introduced in this term structure model may really be unobservable. To handle the unobservable state variables we formulate the model in state space form, a detailed description of which can be found in Harvey (1993) and use the Kalman filter to estimate the parameters (see e.g. Dempster et al. 1999 or Medova et al. 2005). 10.3.2

Pricing Coupon-Bearing Bonds

As sufficient historical data on Euro coupon-bearing Treasury bonds is difficult to obtain we use the zero-coupon yield curve to construct the relevant bonds. Coupons on newlyissued bonds are generally closely related to the ccorresponding spot rate at the time, so the current zero-coupon yield with maturity T is used as a proxy for the coupon rate of a coupon-bearing Treasury bond with maturity T. For example, on scenario v the coupon 10 rate dB2 ðoÞ on a newly issued 10-year Treasury bond at time t
2 will be set equal to the projected 10-year spot rate y2,10 (v) at time t
2. Generally ðT Þ

dBt ðoÞ ¼ yt;T ðoÞ; ðT Þ

ðT Þ

dBt ðoÞ ¼ dbt c ðoÞ;

8t 2 T d ; 8o 2 O;

ð10:41Þ

8t 2 T i ; 8o 2 O;

ð10:42Þ

where  ×  denotes integral part. This ensures that as the yield curve falls, coupons on newly-issued bonds will go down correspondingly and each coupon cash flow will be discounted at the appropriate zero-coupon yield.

DESIGNING MINIMUM GUARANTEED RETURN FUNDS

j

235

The bonds are assumed to pay coupons semi-annually. Since we roll the bonds on an annual basis, a coupon will be received after six months and again after a year just before the bond is sold. This forces us to distinguish between the price at which the bond is sold at rebalancing times and the price at which the new bond is purchased. ðsellÞ Let Pt;BðT Þ denote the selling price of the bond BðT Þ at time t, assuming two coupons ðbuy Þ have now been paid out and the time to maturity is equal to T1, and let Pt;BðT Þ denote the price of a newly issued coupon-bearing Treasury bond with a maturity equal to T. The ‘buy’ bond price at time t is given by T

BtT ðoÞ ¼ F B eðT þbt ct Þyt;T þbt ct ðoÞ T X dBt ðoÞ BT ðst Þyt;ðst Þ ðoÞ F e þ ; 2 b2t c b2t c s¼ þ1; þ1;...;bt cþT 2

2

o 2 O; t 2 T total ;

ð10:43Þ

2

T

where the principal FB of the bond is discounted in the first term and the stream of coupon payments in the second. At rebalancing times t the sell price of the bond is given by T

BtT ðoÞ ¼ F B eðT 1Þyt;T 1 ðoÞ X dBT ðoÞ T t1 F B eðst Þyt;ðst Þ ðoÞ þ 2 1 s¼ ;1;...;T 1

  o 2 O t 2 T d nf0g [ fT g

ð10:44Þ

2

T

with coupon rate dBt1 ðoÞ. The coupon rate is then reset for the newly-issued Treasury bond of the same maturity. We assume that the coupons paid at six months are re-invested in the on-the-run bonds. This gives the following adjustment to the amount held in bond BT at time t.

xt;BT ðoÞ ¼ xt121 ; BT ðoÞ þ

10.4

T T 1 B d ðoÞF B xt121 ; BT ðoÞ 2 t ; buy fPt;BT ðoÞ

t 2 T c ; o 2 O:

ð10:45Þ

HISTORICAL BACKTESTS

We will look at an historical backtest in which statistical models are fitted to data up to a trading time t and scenario trees are generated to some chosen horizon t  T. The optimal root node decisions are then implemented at time t and compared to the historical returns at time t  1. Afterwards the whole procedure is rolled forward for T trading times. Our backtest will involve a telescoping horizon as depicted in Figure 10.4. At each decision time t the parameters of the stochastic processes driving the stock return and the three factors of the term structure model are re-calibrated using historical data up to and including time t and the initial values of the simulated scenarios are given by the actual historical values of the variables at these times. Re-calibrating the simulator

236 j CHAPTER 10 5-Year Scenario Tree 4-Year Scenario Tree 3-Year Scenario Tree 2-Year Scenario Tree 1-Year Scenario Tree Jan 1999

Jan 2000

Jan 2001

Jan 2002

Jan 2003

Jan 2004

FIGURE 10.4 Telescoping horizon backtest schema. TABLE 10.2 Tree Structures for Different Horizon Backtests Jan 1999

6.6.6.6.6  7776

32.4.4.4.4  8192

512.2.2.2.2  8192

Jan 2000 Jan 2001 Jan 2002 Jan 2003

9.9.9.9
6561 20.20.20
8000 88.88
7744 7776

48.6.6.6
10368 80.10.10
8000 256.32
8192 8192

512.2.2.2
4096 768.3.3
6912 1024.8
8192 8192

parameters at each successive initial decision time t captures information in the history of the variables up to that point. Although the optimal second and later-stage decisions of a given problem may be of ‘what-if ’ interest, manager and decision maker focus is on the implementable first-stage decisions which are hedged against the simulated future uncertainties. The reasons for implementing stochastic optimization programs in this way are twofold. Firstly, after one year has passed the actual values of the variables realized may not coincide with any of the values of the variables in the simulated scenarios. In this case the optimal investment policy would be undefined, as the model only has optimal decisions defined for the nodes on the simulated scenarios. Secondly, as one more year has passed new information has become available to re-calibrate the simulator’s parameters. Relying on the original optimal investment strategies will ignore this information. For more on backtesting procedures for stochastic optimization models see Dempster et al. (2003). For our backtests we will use three different tree structures with approximately the same number of scenarios, but with an increasing initial branching factor. We first solve the five-year problem using a 6.6.6.6.6 tree, which gives 7776 scenarios. Then we use 32.4.4.4.4
8192 scenarios and finally the extreme case of 512.2.2.2.2
8192 scenarios. For the subsequent stages of the telescoping horizon backtest we adjust the branching factors in such a way that the total number of scenarios stays as close as possible to the original number of scenarios and the same ratio is maintained. This gives us the tree structures set out in Table 10.2. Historical backtests show how specific models would have performed had they been implemented in practice. The reader is referred to Rietbergen (2005) for the calibrated parameter values employed in these tests. Figures 10.5 to 10.10 show how the various optimal portfolios’ wealth would have evolved historically relative to the barrier. It is also

DESIGNING MINIMUM GUARANTEED RETURN FUNDS Backtest 99-04: 6.6.6.6.6 = 7776 Scenarios Expected Average Shortfall 170 160 150 140 130 120 110 100 90 80 1-Jan-99

1-Jan-00

Barrier

31-Dec-00 31-Dec-01 31-Dec-02 31-Dec-03

EAS

Exp EAS

EAS MC

Exp EAS MC

FIGURE 10.5 Backtest 1999 /2004 using expected average shortfall for the 6.6.6.6.6 tree.

Backtest 99-04: 6.6.6.6.6 = 7776 Scenarios Expected Maximum Shortfall 170 160 150 140 130 120 110 100 90 80 1-Jan-99

1-Jan-00

Barrier

31-Dec-00 31-Dec-01 31-Dec-02 31-Dec-03

EMS

Exp EMS

EMS MC Exp EMS MC

FIGURE 10.6 Backtest 1999 /2004 using expected maximum shortfall for the 6.6.6.6.6 tree.

Backtest 99-04: 32.4.4.4.4 = 8192 Scenarios Expected Average Shortfall 150 140 130 120 110 100 90 80 1-Jan-99

1-Jan-00

Barrier

31-Dec-00 31-Dec-01 31-Dec-02 31-Dec-03

EAS

Exp EAS

EAS MC

Exp EAS MC

FIGURE 10.7 Backtest 1999 /2004 using expected average shortfall for the 32.4.4.4.4 tree.

j

237

238 j CHAPTER 10 Backtest 99-04: 32.4.4.4.4 = 8192 Scenarios Expected Maximum Shortfall 150 140 130 120 110 100 90 80 1-Jan-99

1-Jan-00

Barrier

31-Dec-00 31-Dec-01 31-Dec-02 31-Dec-03

EMS

Exp EMS

EMS MC Exp EMS MC

FIGURE 10.8 Backtest 1999 /2004 using expected maximum shortfall for the 32.4.4.4.4 tree. Backtest 99-04: 512.2.2.2.2 = 8192 Scenarios Expected Average Shortfall 140 130 120 110 100 90 80 1-Jan-99

1-Jan-00

Barrier

31-Dec-00 31-Dec-01 31-Dec-02 31-Dec-03

EAS

Exp EAS

EAS MC Exp EAS MC

FIGURE 10.9 Backtest 1999 /2004 using expected average shortfall for the 512.2.2.2.2 tree. Backtest 99-04: 512.2.2.2.2 = 8192 Scenarios Expected Maximum Shortfall 140 130 120 110 100 90 80 1-Jan-99

1-Jan-00

Barrier

31-Dec-00 31-Dec-01 31-Dec-02 31-Dec-03

EMS

Exp EMS

EMS MC Exp EMS MC

FIGURE 10.10 Backtest 1999 /2004 using expected maximum shortfall for the 512.2.2.2.2 tree.

important to determine how the models performed in-sample on the generated scenario trees and whether or not they had realistic forecasts with regard to future historical returns. To this end one-year-ahead in-sample expectations of portfolio wealth are shown as points in the backtest performance graphs. Implementing the first-stage decisions insample, the portfolio’s wealth is calculated one year later for each scenario in the simulated tree after which an expectation is taken over the scenarios.

DESIGNING MINIMUM GUARANTEED RETURN FUNDS

j

239

From these graphs a first observation is that the risk management monitoring incorporated into the model appears to work well. In all cases the only time portfolio wealth dips below the barrier, if at all, is on 11 September 2001. The initial in-sample wealth overestimation of the models is likely to be due mainly to the short time series available for initial parameter estimation which led to hugely inflated stock return expectations during the equity bubble. However as time progresses and more data points to re-calibrate the model are obtained, the models’ expectations and real-life realizations very closely approximate each other. For reference we have included the performance of the Eurostoxx 50 in Figure 10.11 to indicate the performance of the stock market over the backtesting period. Even though this was a difficult period for the optimal portfolios to generate high historical returns, it provides an excellent demonstration that the risk management incorporated into the models operates effectively. It is in periods of economic downturn that one wants the portfolio returns to survive. Tables 10.3 and 10.4 give the optimal portfolio allocations for the 32.4.4.4.4 tree using the two maximum shortfall objective functions. In both cases we can observe a for the portfolio to move to the safer, shorter-term assets as time progresses. This is naturally built into the model as depicted in Figure 10.3. For the decisions made in January 2002/2003, the portfolio wealth is significantly closer to the barrier for the EMS model than it is for the EMS MC model. This increased risk for the fund is taken into account by the EMS model and results in an investment in safer short-term bonds and a negligible equity component. Whereas the EMS model stays in the one to three year range the EMS MC model invests mainly in bonds with a maturity in the range of three to five years and for both models the portfolio wealth manages to stay above the barrier. From Figures 10.5 to 10.10 it can be observed that in all cases the method with monthly checking outperforms the equivalent method with just annual shortfall checks. Similarly as the initial branching factor is increased, the models’ out-of-sample performance is generally improved. For the 512.2.2.2.2
8192 scenario tree, all four objective functions give optimal portfolio allocations which keep the portfolio wealth above the barrier at all times, but the models with the monthly checking still outperform the others. The more important difference however seems to lie in the deviation of the expected in-sample portfolio’s wealth from the actual historical realization of the portfolio value. Table 10.5 displays this annual deviation averaged over the five rebalances and shows a clear reduction in this deviation for three of the four models as the initial branching factor is increased. Again the model that uses the expected maximum shortfall with monthly checking as its objective function outperforms the rest. Overall the historical backtests have shown that the described stochastic optimization framework carefully considers the risks created by the guarantee. The EMS MC model produces well-diversified portfolios that do not change drastically from one year to the next and results in optimal portfolios which even through a period of economic downturn and uncertainty remained above the barrier.

240 j CHAPTER 10 Backtest 99-04: 512.2.2.2.2 = 8192 Scenarios Expected Maximum Shortfall 160 140 120 100 80 60 1-Jan-99

1-Jan-00

31-Dec-00 31-Dec-01 31-Dec-02 31-Dec-03

Barrier

EMS

EMS MC

Eurostoxx 50

FIGURE 10.11 Comparison of the fund’s portfolio performance to the Eurostoxx 50. TABLE 10.3 Portfolio Allocation Expected Maximum Shortfall Using the 32.4.4.4.4 Tree

Jan 99 Jan 00 Jan 01 Jan 02 Jan 03

1y

2y

3y

4y

5y

10y

30y

Stock

0 0 0.04 0.08 0.92

0 0 0 0.16 0

0 0 0 0.74 0

0 0 0 0 0

0 0 0 0 0

0.23 0 0.39 0 0.07

0.45 0.37 0.53 0 0

0.32 0.63 0.40 0.01 0.01

TABLE 10.4 Portfolio Allocation Expected Maximum Shortfall with Monthly Checking Using the 32.4.4.4.4 Tree

Jan 99 Jan 00 Jan 01 Jan 02 Jan 03

1y

2y

3y

4y

5y

10y

30y

Stock

0 0 0 0 0

0 0 0 0 0

0 0 0 0 0.78

0 0 0 0.47 0.22

0.49 0.25 0.49 0.44 0

0.27 0.38 0.15 0 0

0 0 0 0 0

0.24 0.36 0.36 0.10 0.01

TABLE 10.5 Average Annual Deviation

6.6.6.6.6 32.4.4.4.4 512.2.2.2.2

10.5

EAS

EAS MC

EMS

9.87) 10.06) 10.22)

13.21) 9.41) 8.78)

9.86) 9.84) 7.78)

EMS MC 10.77) 7.78) 6.86)

ROBUSTNESS OF BACKTEST RESULTS

Empirical equity returns are now well known not to be normally distributed but rather to exhibit complex behaviour including fat tails. To investigate how the EMS MC model performs with more realistic asset return distributions we report in this section experiments using a geometric Brownian motion with Poisson jumps to model equity returns. The stock price process S is now assumed to follow

DESIGNING MINIMUM GUARANTEED RETURN FUNDS

j

241

130 120 110 100 90 80 1-Jan-99

1-Jan-00

31-Dec-00

Barrier

31-Dec-01

EMS MC

31-Dec-02

31-Dec-03

Exp EMS MC

FIGURE 10.12 Expected maximum shortfall with monthly checking using the 512.2.2.2.2 tree for the GBM with jumps equity index process. TABLE 10.6 Portfolio Allocation Expected Maximum Shortfall with Monthly Checking Using the 512.2.2.2.2 Tree

Jan 99 Jan 00 Jan 01 Jan 02 Jan 03

1y

2y

3y

4y

5y

10y

30y

Stock

0 0 0 0 0

0 0 0 0 0

0 0 0 0 0.05

0 0 0 0 0

0.69 0.63 0.37 0.90 0.94

0.13 0 0.44 0 0

0 0 0 0 0

0.18 0.37 0.19 0.10 0.01

TABLE 10.7 Portfolio Allocation Expected Maximum Shortfall with Monthly Checking Using the 512.2.2.2.2 Tree for the GBM with Poisson Jumps Equity Index Process

Jan 99 Jan 00 Jan 01 Jan 02 Jan 03

1y

2y

3y

4y

5y

10y

30y

Stock

0 0 0 0 0

0 0 0 0 0

0 0 0 0 0

0 0 0 0 0

0.12 0 0.43 0.70 0.04

0.77 0.86 0.56 0.11 0.81

0 0 0 0 0

0.11 0.14 0.01 0.19 0.15

dSt St

S

~ t þ Jt dNt ; ~S dt þ s ~ S dW ¼m

ð10:46Þ

where N is an independent Poisson process with intensity l and the jump saltus J at Poisson epochs is an independent normal random variable. As the EMS MC model and the 512.2.2.2.2 tree provided the best results with Gaussian returns the backtest is repeated for this model and treesize. Figure 10.12 gives the historical backtest results and Tables 10.6 and 10.7 represent the allocations for the 512.2.2.2.2 tests with the EMS MC model for the original GBM process and the GBM with Poisson jumps process respectively. The main difference in the two tables is that the

242 j CHAPTER 10

investment in equity is substantially lower initially when the equity index volatility is high (going down to 0.1) when the volatility is 28) in 2001), but then increases as the volatility comes down to 23) in 2003. This is born out by Figure 10.12 which shows much more realistic in-sample one-year-ahead portfolio wealth predictions (cf. Figure 10.10) and a 140 basis point increase in terminal historical fund return over the Gaussian model. These phenomena are the result of the calibration of the normal jump saltus distributions to have negative means and hence more downward than upwards jumps resulting in downwardly skewed equity index return distributions, but with the same compensated drift as in the GBM case. As a consequence the optimal portfolios are more sensitive to equity variation and take benefit from its lower predicted values in the last two years. Although much more complex equity return processes are possible, these results show that the historical backtest performance of the EMS MC model is only improved in the presence of downwardly skewed asset equity return distributions possessing fat tails due to jumps.

10.6

CONCLUSIONS

This chapter has focused on the design of funds to support investment products which give a minimum guaranteed return. We have concentrated here on the design of the liability side of the fund, paying particular attention to the pricing of bonds using a threefactor term structure model with reliable results for long-term as well as the short-term yields. Several objective functions for the stochastic optimization of portfolios have been constructed using expected average shortfall and expected maximum shortfall risk measures to combine risk management with strategic asset allocation. We also introduced the concept of monthly shortfall checking which improved the historical backtesting results considerably. In addition to the standard GBM model for equity returns we reported experiments using a GBM model with Poisson jumps to create downwardly skewed fat tailed equity index return distributions. The EMS MC model responded well with more realistic expected portfolio wealth predictions and the historical fund portfolio wealth staying significantly above the barrier at all times. The models of this chapter have been extended in practice to open ended funds which allow for contributions throughout the lifetime of the corresponding investment products. In total funds of the order of 10 billion euros have been managed with these extended models. In future research we hope to examine open multi-link pension funds constructed using several unit linked funds of varying risk aversion in order to allow the application of individual risk management to each client’s portfolio.

ACKNOWLEDGEMENTS The authors wish to thank Drs. Yee Sook Yong and Ning Zhang for their able assistance with computational matters. We also acknowledge the helpful comments of two anonymous referees which led to material improvements in the chapter.

DESIGNING MINIMUM GUARANTEED RETURN FUNDS

j

243

REFERENCES Andersen, T.G. and Lund, J., Estimating continuous-time stochastic volatility models of the shortterm interest rate. J. Econometrics, 1997, 77(2), 343/377. Campbell, R., The economic factor model: Theory and applications. Lehman Brothers Presentation to Europlus Research and Management Dublin, 31 March 2000. Consigli, G. and Dempster, M.A.H., The CALM stochastic programming model for dynamic assetliability management. In Worldwide Asset and Liability Modeling, edited by J.M. Mulvey and W.T. Ziemba, pp. 464 /500, 1998 (Cambridge University Press: Cambridge). Consiglio, A., Cocco, F. and Zenios, S.A., The PROMETEIA model for managing insurance policies with guarantees. In Handbook of Asset and Liability Management, edited by S. Zenios and W.T. Ziemba, Vol. 2, 2007 (North-Holland: Oxford); in press. Cox, J.C., Ingersoll, J.E. and Ross, S.A., A theory of the term structure of interest rates. Econometrica, 1985, 53, 385/407. Dempster, M.A.H., Germano, M., Medova, E.A. and Villaverde, M., Global asset liability management. Br. Actuarial J., 2003, 9(1), 137 /216. Dempster, M.A.H., Hicks-Pedron, N., Medova, E.A., Scott, J.E. and Sembos, A., Planning logistics operations in the oil industry. J. Operat. Res. Soc, 2000, 51, 1271 /1288. Dempster, M.A.H., Jones, C.M., Khokhar, S.Q. and Hong, G.S-S., Implementation of a model of the stochastic behaviour of commodity prices. Report to Rio Tinto, Centre for Financial Research, Judge Institute of Management, University of Cambridge, 1999 . Harvey, A.C., Time Series Models, 1993 (Harvester Wheatsheaf: Hemel Hempstead). Hertzog, F., Dondi, G., Keel, S., Schumani, L.M. and Geering, H.P., Solving ALM problems via sequential stochastic programming. Quant. Finance, 2007, 7(2), 231 /244. Hull, J.C., Options, Futures and Other Derivatives, 1997 (Prentice Hall: Upper Saddle River, NJ). Markowitz, H.M., Portfolio Selection, 1959 (Wiley: New York). Medova, E.A., Rietbergen, M.I., Villaverde, M. and Yong, Y.S., Modelling the long-term dynamics of the yield curve. Working Paper, 2005 (Centre for Financial Research, Judge Business School, University of Cambridge). Nelson, C.R. and Siegel, A.F., Parsimonious modeling of yield curves. J. Bus, 1987, 60, 473/489. Rietbergen, M.I., Long term asset liability management for minimum guaranteed return funds. PhD Dissertation, Centre for Financial Research, Judge Business School, University of Cambridge, 2005. Vasicek, O., An equilibrium characterization of the term structure. J. Financ. Econ, 1977, 5(2), 177/188.

PART

Portfolio Construction and Risk Management

2

11

CHAPTER

DC Pension Fund Benchmarking with Fixed-Mix Portfolio Optimization M. A. H. DEMPSTER, M. GERMANO, E. A. MEDOVA, M. I. RIETBERGEN, F. SANDRINI, M. SCROWSTON and N. ZHANG CONTENTS 11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Current Market Practice . . . . . . . . . . . . . . . . 11.3 Optimal Benchmark Definition for DC Funds. 11.4 Fund Model . . . . . . . . . . . . . . . . . . . . . . . . 11.5 Nominal Guarantee Results . . . . . . . . . . . . . 11.6 Inflation-Linked Guarantee Results . . . . . . . . 11.7 Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11.1

C

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

247 248 249 250 252 254 257 258

INTRODUCTION

pension schemes have recently found themselves in hot water. Accounting practices that led to over-exposure to equity markets, increases in longevity of the scheme participants and low interest rates have all contributed to the majority of schemes in the EU and the U.K. finding themselves underfunded. In essence, a DB scheme promises to pay its participants an annuity at retirement that gives them a pension equal to a proportion of their final salary (the proportion depending on the number of years of service). Therefore the responsibility to meet these promises (liabilities) rests firmly with the scheme’s trustees and ultimately with the corporate sponsor. The management of these corporate schemes was greatly affected in the past by quarterly earnings reports which directly impacted stock prices in the quest for ‘shareholder value.’ ORPORATE SPONSORED DEFINED BENEFIT (DB)

247

248 j CHAPTER 11

Consequently DB scheme sponsors resorted to a management style that was able to keep the liabilities, if not off the balance sheet, then at least to a minimum. One sanctioned tactic that achieved these aims was the ability to discount liabilities by the expected return of the constituent asset classes of the fund. In other words, by holding a large part of the fund in equities, the liabilities could be discounted away at over 10) p.a. The recent performance of the equity markets and the perception of equity as a long-horizon asset class assisted in justifying this asset-mix in the eyes of the scheme’s trustees. However, with the collapse of the equity-market bubble in 2001, many funds found their schemes grossly underfunded and were forced to crystallize their losses by panicked trustees. Consequent tightening of the regulations has made the situation even worse (e.g. all discounting must be done by the much lower AA credit quality bond yield rates in the UK FRS17 standard). As a result many DB schemes have closed and are now being replaced with defined contribution (DC) schemes.1 In this world of corporate sponsored DC pension schemes the liability is separated from the sponsor and the market risk is placed on the shoulders of the participants. The scheme is likely to be overseen by an investment consultant and if the scheme invests in funds that perform badly over time a decision may be made by the consultant to move the capital to another fund. However, any losses to the fund will be borne by the participants in the scheme and not by the corporate sponsor. Since at retirement date scheme participants will wish to either purchase an annuity or invest their fund payout in a self-managed portfolio, an obvious need arises in the market place for real return guaranteed schemes which are similar to those often found in life insurance policies. These guarantees will typically involve inflation protection plus some element of capital growth, for example, inflation rate plus 1) per annum. From the DC fund manager’s viewpoint provision of the relevant guarantee requires very tight risk, control, as the recent difficulties at Equitable Life so graphically illustrate. The question addressed in this chapter is how consultants or DC fund-managers can come to a sensible definition of an easily understandable liability-related benchmark against which the overall fund performance for a DC scheme can be measured. Performance of both fund and benchmark must be expressed to fund participants in easy-to-understand concepts such as probability of achieving some target wealth level above the scheme guarantee—a measure easily derived from the solutions of the models discussed in this chapter.

11.2

CURRENT MARKET PRACTICE

Currently, DC pension funds are market-benchmarked against either a fixed-mix of defined asset classes (total return bond and equity indices) or against some average performance of their peers. The benchmark is not defined in terms of the pension liability and investment is not liability-driven. The standard definitions of investment risk—standard deviation, semi-variance and downside risk—do not convey information regarding the probability of missing the scheme participants’ investment goals and obligations. 1

DC pension scheme participants typically make a lump sum initial payment and regular contributions to the pension fund which are employer matched.

DC PENSION FUND BENCHMARKING WITH FIXED-MIX PORTFOLIO

j

249

For example, the macro-asset benchmark may be defined as 20) of certain equity indices and 80) of particular bond indices, but may not reflect the risk that the scheme participants are willing to take in order to attain a specific substitution rate between their final salary and pension income (pension earnings/final salary).

11.3

OPTIMAL BENCHMARK DEFINITION FOR DC FUNDS

In line with current market practice, we wish to find a definition of a fixed asset mix benchmark similar to that given in the example above but with an asset mix that optimizes returns against user-defined risk preferences. Specifically, for participants that are willing to take a certain amount of risk in order to aim for a given substitution rate between final salary and pension, we should be able to ‘tune’ the asset mix in an optimal way to reflect the participants desire to reach this substitution rate. The risk could then be defined as the probability of not reaching that substitution rate. In contrast to dynamic multi-stage portfolio optimization, where the asset-mix is changed dynamically over time to reflect changing attitudes to risk as well as market performance (dynamic utility), a fixed-mix rebalance strategy benchmark in some sense reflects an average of this dynamic utility over the fund horizon. For such a strategy the realized portfolio at each decision stage is rebalanced back to a fixed set of portfolio weights. In practice for DC pension schemes we want the returns to be in line with salary inflation in the sense that the required substitution rate is reached with a given probability. The solution to this problem will entail solving a fixed-mix dynamic stochastic programme that reflects the long run utility of the scheme participants. In general multi-period dynamic stochastic optimization will be more appropriate for longterm investors. Single-period models construct optimal portfolios that remain unchanged over the planning horizon while fixed mix rebalance strategies fail to consider possible investment opportunities that might arise due to market conditions over the course of the investment horizon. Dynamic stochastic programmes on the other hand capture optimally an investment policy in the face of the uncertainty about the future given by a set of scenarios. Carino and Turner (1998) compare a multi-period stochastic programming approach to a fixed-mix strategy employing traditional mean-variance efficient portfolios. Taking a portfolio from the mean-variance efficient frontier, it is assumed that the allocations are rebalanced back to that mix at each decision stage. They also highlight the inability of the mean-variance optimization to deal with derivatives such as options due to the skewness of the resulting return distributions not being taken into account. The objective function of the stochastic programme is given by maximizing expected wealth less a measure of risk given by a convex cost function. The stochastic programming approach was found to dominate fixed-mix in the sense that for any given fixed-mix rebalance strategy, there is a strategy that has either the same expected wealth and lower shortfall cost, or the same shortfall cost and higher expected wealth. Similar results were found by Hicks-Pedro´n (1998) who also showed the superiority in terms of final Sharpe ratio of both methods to the constant proportion portfolio insurance (CPPI) strategy over long horizons. Fleten et al. (2002) compare the performance of four-stage stochastic models to fixedmix strategies of in- and out-of-sample, using a set of 200 flat scenarios to obtain the

250 j CHAPTER 11

out-of-sample results. They show that the dynamic stochastic programming solutions dominate the fixed-mix solutions both in- and out-of-sample, although to a lesser extent out-of-sample. This is due to the ability of the stochastic programming model to adapt to the information in the scenario tree in-sample, although they do allow the fixed-mix solution to change every year once new information has become available, making this sub-optimal strategy inherently more dynamic. Mulvey et al. (2003) compare buy-and-hold portfolios to fixed-mix portfolios over a ten-year period, showing that in terms of expected return versus return standard deviation, the fixed-mix strategy generates a superior efficient frontier, where the excess returns are due to portfolio rebalancing. Dempster et al. (2007a) discuss the theoretical cause of this effect (and the historical development of its understanding) under the very general assumption of stationary ergodic returns. A similar result was found in Mulvey et al. (2004) with respect to including alternative investments into the portfolio. In particular they looked at the use of the Mt. Lucas Management index in multi-period fixed-mix strategies. A multi-period optimization will not only identify these gains but also take advantage of volatility by suggesting solutions that are optimal in alternative market scenarios. In Mulvey et al. (2007) the positive long term performance effects of new asset classes, leverage and various overlay strategies are demonstrated for both fixed-mix and dynamically optimized strategies.

11.4

FUND MODEL

The dynamic optimal portfolio construction problem for a DC fund with a performance guarantee is modelled here at the strategic level with annual rebalancing. The objective is to maximize the expected sum of accumulated wealth while keeping the expected maximum shortfall of the portfolio relative to the guarantee over the 5 year planning horizon as small as possible. A complete description of the dynamic stochastic programming model can be found in Dempster et al. (2006). In the fixed-mix model the portfolio is rebalanced to fixed proportions at all future decision nodes, but not at the intermediate time stages used for shortfall checking. This results in annual rebalancing while keeping the risk management function monthly, leading to the objective function for both problems as n

max

( o ð1  bÞ

þ  xt;a ðoÞ;xt;a ðoÞ;xt;a ðoÞ: a2A;o2O;t2T d [fT g

X o2O

b

X

pðoÞ

X

! Wt ðoÞ

t2T d [fT g

pðoÞ maxt2T total ht ðoÞ

!) ;

ð11:1Þ

o2O

where “ “ “

p(v) denotes the probability of scenario v in V—here p(v):  1uN with N scenarios, Wt (v) denotes the portfolio wealth at time t
Ttotal in scenario v, ht (v) denotes the shortfall relative to the barrier at time t in scenario v.

DC PENSION FUND BENCHMARKING WITH FIXED-MIX PORTFOLIO

j

251

For the nominal or fixed guarantee, the barrier at time t in scenario v, below which the fund will be unable to meet the guarantee, is given by T

T

LtF ðoÞ ¼ W0 ð1 þ GÞ Zt ðoÞ ¼ W0 ð1 þ GÞ eyt;T ðoÞðT tÞ ;

ð11:2Þ

where “ “

G denotes the annual nominal guarantee Zt (v) denotes the zero-coupon Treasury bond price at time t in scenario v.

For simplicity we model closed-end funds here, but see Dempster et al. (2006, 2007) and Rietbergen (2005) for the treatment of contributions. We employ a five-period (stage) model with a total of 8192 scenarios to obtain the solutions for the dynamic optimization and fixed mix approaches.2 For this chapter, five different experiments were run on a five-year closed-end fund with a minimum nominal guarantee of 2) and an initial wealth of 100 using a 512.2.2.2.2 tree.3 (See Dempster et al. (2006, 2007) for more details on this problem). The parameter of risk aversion b is set to 0.99 and the parameter values used were estimated over the period June 1997 December 2002. The Pioneer CASM simulator was used to generate the problem data at monthly intervals. The five experiments run were as follows. /

“

“ “ “

“

Experiment 1: No fixed-mix constraints. Objective function: fund wealth less expected maximum shortfall with monthly checking. Experiment 2: Arbitrary fixed-mix: 30) equity and 10) in each of the bonds. Experiment 3: The fixed-mix is set equal to the root node decision of Experiment 1. Experiment 4: The fixed-mix is set equal to the root node decision of Experiment 1 but only applied after the first stage. The root node decision is optimized. Experiment 5: The fixed-mix is determined optimally.

Experiments 2 4 with fixed-mixed constraints are ‘fixed fixed-mix’ problems in which the fixed-mix is specified in advance in order to keep the optimization problem convex. Finally Experiment 5 uses fixed-mix constraints without fixing them in advance. This renders the optimization problem non-convex so that a global optimization technique needs to be used. In preliminary experiments we found that although the resulting unconstrained problems are multi-extremal they are ‘near-convex’ and can be globally optimized by a search routine followed by a local convex optimizer. For this purpose we used Powell’s (1964) algorithm followed by the SNOPT solver. Function evaluations involving all fixed-mix rebalances were evaluated by linear programming using CPLEX. This method is described in detail in Scott (2002). As the fixed-mix policy remains the same at all rebalances, theoretically there is no reason to have a scenario tree which branches more than once at the beginning of the first /

2

In practice 10 and 15 year horizons have also been employed. Assets employed are Eurobonds of 1, 2, 3, 4, 5, 10 and 30 year maturities and equity represented by the Eurostock 50 index.

3

252 j CHAPTER 11

year. A simple fan tree structure would be perfectly adequate as the fixed-mix approach is unable to exploit the perfect foresight implied after the first stage in this tree. However for comparison reasons we use the same tree for both the dynamic stochastic programme (Experiment 1) and the fixed-mix approach (Experiments 2 5). /

11.5

NOMINAL GUARANTEE RESULTS

Table 11.1 shows the expected terminal wealth and expected maximum shortfall for the five experiments. As expected, Experiment 1 with no fixed-mix constraints results in the highest expected terminal wealth and lowest expected maximum shortfall. Whereas Experiments 2 and 3 underperform, Experiment 3 in which the initial root node solution of Experiment 1 is used as the fixed-mix is a significant improvement on Experiment 2 (arbitrary fixed-mix) and might serve as an appropriate benchmark. Experiment 4 resulted in a comparable expected terminal wealth to Experiment 3, but the expected maximum shortfall is now an order of magnitude smaller. Finally in Experiment 5 global optimization was used which correctly resulted in an improvement relative to Experiment 3 in both the expected terminal wealth and the expected maximum shortfall. Table 11.2 shows the optimal root node decisions for all five experiments. With the equity market performing badly and declining interest rates over the 1997 2002 period, we see a heavy reliance on bonds in all portfolios. Figures 11.1 and 11.2 show the efficient frontiers for the dynamic stochastic programme and the fixed-mix solution, where the risk measure is given by expected maximum shortfall. Figure 11.1 shows that the dynamic stochastic programme generates a much bigger range of possible risk return trade-offs and even if we limit the range of risk parameters to that given for the fixed-mix experiments as in Figure 11.2, we see that the DSP problems clearly outperform the fixed-mix problems. /

TABLE 11.1 Expected Terminal Wealth and Maximum Shortfall for the Nominal Guarantee

Experiment Experiment Experiment Experiment Experiment

1 2 3 4 5

Expected terminal wealth

Expected maximum shortfall

126.86 105.58 120.69 119.11 122.38

8.47 E-08 14.43 0.133 0.014 0.122

TABLE 11.2 Root Node Solutions for the Nominal Guarantee

Exp Exp Exp Exp Exp

1 2 3 4 5

1y

2y

3y

4y

5y

10y

30y

Stock

0 0.10 0 0 0

0 0.10 0 0 0

0.97 0.10 0.97 0.06 0.96

0 0.10 0 0.94 0.04

0 0.10 0 0 0

0 0.10 0 0 0

0.02 0.10 0.02 0 0

0.01 0.30 0.01 0 0

DC PENSION FUND BENCHMARKING WITH FIXED-MIX PORTFOLIO

j

Expected Terminal Wealth

Efficient Frontier 140 138 136 134 132 130 128 126 124 122 120

DSP Fixed-mix 0

15 10 Expected Maximum Shortfall

5

20

FIGURE 11.1 Efficient frontier for the nominal guarantee.

Efficient Frontier Expected Terminal Wealth

130 129 128 127 126 125 124 123

DSP

122

Fixed-mix

121 0

0.05

0.1 0.15 0.2 0.25 Expected Maximum Shortfall

0.3

0.35

FIGURE 11.2 Efficient frontier for the nominal guarantee.

0

0

0

21 0 21 5

20

19

18

0 17

0

0

0

0

16

15

14

0

13

12

0 11

10

0

Probability

No Fixed-Mix Restrictions 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 Terminal Wealth

FIGURE 11.3 Terminal wealth distribution for optimal dynamic stochastic policy of Experiment 1.

253

254 j CHAPTER 11 Fixed-Mix 0.3

Probability

0.25 0.2 0.15 0.1 0.05 0 100

105

110

115

120 125 130 135 Terminal Wealth

140

145

150

FIGURE 11.4 Terminal wealth distribution for the optimal fixed-mix policy of Experiment 5.

We also considered the distribution of the terminal wealth as shown in Figures 11.3 and 11.4. From Figure 11.3 we observe a highly skewed terminal wealth distribution for Experiment 1 with most of the weight just above the guaranteed wealth of 110.408. The effect of dynamic allocation is to alter the overall probability distribution of the final wealth. Carino and Turner (1998) also note this result in their experiments. For Experiments 2 and 3 in which there is no direct penalty in the optimization problem for shortfall, we see the more traditional bell-shaped distribution. Using the initial root node solution of Experiment 1 as the fixed-mix portfolio results in a distribution with a higher mean and lower standard deviation (the standard deviation drops from 20.51 to 6.43). In Experiment 4 we see an increase again in the probability of the terminal wealth ending up just above the minimum guarantee of 110 as the optimization problem has flexibility at the initial stage. The standard deviation is further reduced in this experiment to 4.04. The mean and standard deviation of Experiment 5 is comparable to that of Experiment 3, which is as expected since the portfolio allocations of the two experiments are closely related.

11.6

INFLATION-LINKED GUARANTEE RESULTS

In the case of an inflation-indexed guarantee the final guarantee at time T is given by

W0

T Y

1 þ isðmÞ ðoÞ ;

ð11:3Þ

s¼1=12

where isðmÞ ðoÞ represents the monthly inflation rate at time s in scenario v. However, unlike the nominal guarantee, at time t B T the final inflation-linked guarantee is still unknown. We propose to approximate the final guarantee by using the inflation rates which are known at time t, combined with the expected inflation at time t for the period ½t þ ð1=12Þ; T .

DC PENSION FUND BENCHMARKING WITH FIXED-MIX PORTFOLIO

j

255

The inflation-indexed barrier at time t is then given by

0 10 1 t T Y Y



1 þ isðmÞ ðoÞ A@ 1 þ isðmÞ ðoÞ AZt ðoÞ LtI ðoÞ ¼ W0 @ s¼121

s¼tþ121

s¼121

s¼tþ121

0 10 1 t T Y Y



1 þ isðmÞ ðoÞ A@ 1 þ isðmÞ ðoÞ Aeyt;T ðoÞðT tÞ : ¼ W0 @

ð11:4Þ

In general the expected terminal wealth is higher for the inflation-linked barrier, but we also see an increase in the expected maximum shortfall (see Table 11.3). This reflects the increased uncertainty related to the inflation-linked guarantee which also forces us to increase the exposure to more risky assets. With an inflation-linked guarantee the final guarantee is only known for certain at the end of the investment horizon. Relative to the nominal guarantee results of Table 11.2, Table 11.4 shows that the initial portfolio allocations for the inflation-linked guarantee are more focused on long-term bonds.

TABLE 11.3 Expected Terminal Wealth and Maximum Shortfall for the Inflation-Linked Guarantee

Experiment Experiment Experiment Experiment Experiment

1 2 3 4 5

Expected terminal wealth

Expected maximum shortfall

129.88 122.81 129.34 129.54 128.23

0.780 13.60 1.580 1.563 1.456

TABLE 11.4 Root Node Solution for the Inflation-Linked Guarantee

Exp Exp Exp Exp Exp

1 2 3 4 5

1y

2y

3y

4y

5y

10y

30y

Stock

0 0.10 0 0 0

0 0.10 0 0 0

0 0.10 0 0 0

0 0.10 0 0 0

0.77 0.10 0.77 0.88 0.94

0.23 0.10 0.23 0.12 0.06

0 0.10 0 0 0

0 0.30 0 0 0

256 j CHAPTER 11

As in Figure 11.3, Figure 11.5 shows that there is a noticeable pattern of asymmetry in the final wealth outcomes for the dynamic stochastic programme of Experiment 1. However the skewness is not so marked. This is due to the fact that for the inflation-linked guarantee problems inflation rates differ on each scenario and the final guarantee is scenario dependent which results in a different value of the barrier being pursued along each scenario. This symmetrizing effect is even more marked for the inflation-linked guarantee as shown in Figure 11.6 (cf. Figure 11.4). Figure 11.7 shows for the inflation-linked guarantee problem similar out-performance of DSP relative to the optimal fixed-mix policy as in Figure 11.2.

0.35 0.3

Probability

0.25 0.2 0.15 0.1 0.05 0 100 105 110 115 120 125 130 135 140 145 150 155 160 165 170 175 Terminal Wealth

FIGURE 11.5 Terminal wealth distribution for Experiment 1 for the inflation-linked guarantee.

0.45 0.4 0.35 Probability

0.3 0.25 0.2 0.15 0.1 0.05 0 110

115

120

125

130

135

140

Terminal Wealth

FIGURE 11.6 Terminal wealth distribution for Experiment 5 for the inflation-linked guarantee.

DC PENSION FUND BENCHMARKING WITH FIXED-MIX PORTFOLIO

j

257

Efficient Frontier 136

Expected Terminal Wealth

135 134 133 132 131 130 129 DSP

128

Fixed-mix

127 0

0.5

1 1.5 2 Expected Maximum Shortfall

2.5

3

FIGURE 11.7 Efficient frontier for the inflation-linked guarantee.

11.7

CONCLUSION

In this chapter we have compared the performance of two alternative versions of a dynamic portfolio management model for a DC pension scheme which accounts for the liabilities arising from a guaranteed fund return. The results show that a fixed-mixed rebalance policy can be used as a benchmark for the dynamic stochastic programming optimal solution with less complexity and lower computational cost. Whereas the riskreturn trade-off for a fixed-mix portfolio rebalancing strategy is constant over the planning horizon, for the dynamic stochastic programming solutions portfolio allocations shift to less volatile assets as the excess over the liability barrier is reduced. The resulting guarantee shortfall risk for the easy-to-explain fixed-mix portfolio rebalancing strategy is therefore higher and its portfolio returns are lower than those of the dynamic optimal policy. On a percentage basis however these differences are sufficiently small to be able to use the easierto-compute fixed-mix results as a conservative performance benchmark for both insample (model) and actual out-of-sample fund performance. For out-of-sample historical backtests of optimal dynamic stochastic programming solutions for these and related problems the reader is referred to Dempster et al. (2006, 2007b). Perhaps the easiest way to explain both benchmark and actual fund performances to DC pension scheme participants is to give probabilities of achieving (expected) guaranteed payouts and more. These are easily estimated a priori by scenario counts in both models considered in this paper for fund design and risk management. It is of course also possible to link final fund payouts to annuity costs and substitution rates with the corresponding probability estimates.

258 j CHAPTER 11

REFERENCES Carino, D.R. and Turner, A.L., Multiperiod asset allocation with derivative assets. In Worldwide Asset and Liability Modeling, edited by W.T. Ziemba and J.M. Mulvey, pp. 182/204, 1998 (Cambridge University Press: Cambridge). Dempster, M.A.H., Germano, M., Medova, E.A., Rietbergen, M.I., Sandrini, F. and Scrowston, M., Managing guarantees. J. Portfolio Manag., 2006, 32(2), 51 /61. Dempster, M.A.H., Evstigneev, I.V. and Schenk-Hoppe´, K.R., Volatility-induced financial growth Quant. Finance, 2007a, 7(2), 151 /160. Dempster, M.A.H., Germano, M., Medova, E.A., Rietbergen, M.I., Sandrini, F. and Scrowston, M., Designing minimum guaranteed return funds. Quant. Finance, 2007b, 7(2), 245 /256. Fleten, S.E., Hoyland, K. and Wallace, S.W., The performance of stochastic dynamic and fixed mix portfolio models. Eur. J. Oper. Res., 2002, 140(1), 37 /49. Hicks-Pedro´n, N., Model-based asset management: A comparative study. PhD dissertation, Centre for Financial Research, Judge Institute of Management, University of Cambridge, 1998. Mulvey, J.M., Pauling, W.R. and Madey, R.E., Advantages of multiperiod portfolio models. J. Portfolio Manag., 2003, 29(2), 35 /45. Mulvey, J.M., Kaul, S.S.N. and Simsek, K.D., Evaluating a trend-following commodity index for multi-period asset allocation J. Alternat. Invest., 2004, 7(1), 54 /69. Mulvey, J.M., Ural, C. and Zhang, Z., Improving performance for long-term investors: Wide diversification, leverage and overlay strategies. Quant. Finance, 2007, 7(2), 175 /187. Powell, M.J.D., An efficient method for finding the minimum of a function of several variables without calculating derivatives. Comput. J., 1964, 7, 303 /307. Rietbergen, M.I., Long term asset liability management for minimum guaranteed return funds. PhD dissertation, Centre for Financial Research, Judge Business School, University of Cambridge, 2005. Scott, J.E., Modelling and solution of large-scale stochastic programmes. PhD dissertation, Centre for Financial Research, Judge Institute of Management, University of Cambridge, 2002.

















12

CHAPTER

Coherent Measures of Risk in Everyday Market Practice CARLO ACERBI CONTENTS 12.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2 Coherency Axioms and the Shortcomings of VaR 12.3 The Objectivist Paradigm . . . . . . . . . . . . . . . . . 12.4 Estimability. . . . . . . . . . . . . . . . . . . . . . . . . . . 12.5 The Diversification Principle Revisited . . . . . . . . 12.6 Spectral Measures of Risk . . . . . . . . . . . . . . . . . 12.7 Estimators of Spectral Measures. . . . . . . . . . . . . 12.8 Optimization of CRMs: Exploiting Convexity . . . 12.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

12.1

T

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

259 259 262 263 265 266 267 268 269 270

MOTIVATION

of the recent (sometimes very technical) literature on Coherent Risk Measures (CRMs). Our purpose is to review the theory of CRMs from the perspective of practical risk management applications. We have tried to single out those results from the theory which help to understand which CRMs can today be considered as realistic candidate alternatives to Value at Risk (VaR) in financial risk management practice. This has also been the spirit of the author’s research line in recent years Acerbi 2002, Acerbi and Simonetti 2002, Acerbi and Tasche 2002a, b (see Acerbi 2003 for a review).

12.2

HIS CHAPTER IS A GUIDED TOUR

COHERENCY AXIOMS AND THE SHORTCOMINGS OF VAR

In 1997, a seminal paper by Artzner et al. (1997, 1999) introduced the concept of a Coherent Measure of Risk, by imposing, via an axiomatic framework, specific mathematical conditions which enforce some basic principles that a sensible risk measure should always satisfy. This cornerstone of financial mathematics was welcomed by many as 259

260 j CHAPTER 12

the first serious attempt to give a precise definition of financial risk itself, via a deductive approach. Among the four celebrated axioms of coherency, a special role has always been played by the so-called subadditivity axiom rðX þ Y Þ  rðXÞ þ rðY Þ;

ð12:1Þ

where r( ×) represents a measure of risk acting on portfolios’ profit
loss r.v.’s X, Y over a chosen time horizon. The reason why this condition has been long debated is probably due to the fact that VaR—the most popular risk measure for capital adequacy purposes—turned out to be not subadditive and consequently not coherent. As a matter of fact, since inception, the development of the theory of CRMs has run in parallel with the debate on whether and how VaR should be abandoned by the risk management community. The subadditivity axiom encodes the risk diversification principle. The quantity /

HðX; Y ; rÞ ¼ rðXÞ þ rðY Þ  rðX þ Y Þ

ð12:2Þ

is the hedging benefit or, in capital adequacy terms, the capital relief associated with the merging of portfolios X and Y. This quantity will be larger when the two portfolios contain many bets on the same risk driver, but of opposite direction, which therefore hedge each other in the merged portfolio. It will be zero in the limiting case when the two portfolios bet on the same directional move of every common risk factor. But the problem with nonsubadditive risk measures such as VaR is that there happen to be cases in which the hedging benefit turns out to be negative, which is simply nonsensical from a risk
theoretical perspective. Specific examples of subadditivity violations of VaR are available in the literature (Artzner et al. 1999; Acerbi and Tasche 2002a), which typically involve discrete distributions. It may be surprising to know however that examples of subadditivity violations of VaR can also be built with very inoffensive distributions. For instance, a subadditivity violation of VaR with confidence level p  10 2 occurs with two portfolios X and Y which are identically distributed as standard univariate Gaussian N ð0; 1Þ, but are dependent through the copula function

/

Cðx; yÞ ¼ dðx  yÞ 1fx > q;y > qg þ dðq  x  yÞ 1fx < q;y < qg

ð12:3Þ

with q  p
. For instance, choosing
 10 4 one finds HðX; Y ; VaRp Þ ’ 2:4 þ 2:4  5:2 ¼ 0:4:

ð12:4Þ

This may be surprising for the many people who have always been convinced that VaR is subadditive in the presence of ‘normal distributions.’ This is in fact true, provided that the full joint distribution is a multivariate Gaussian, but it is false, as in this case, when only the two marginal distributions are Gaussian. This leads to the conclusion that it is never sufficient to study the marginals to ward off a VaR violation of subadditivity, because the trigger of such events is a subtler copula property.

COHERENT MEASURES OF RISK IN EVERYDAY MARKET PRACTICE

j

261

70 60 50 40 30 20 10 0 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

FIGURE 12.1 The one-dimensional non-convex VaR5)(Pl) risk surface (dotted line) and the convex ES5)(Pl) risk surface (solid line) for a specific portfolio Pl l
[0, 1]. From Acerbi (2003), Example 2.15.

Other examples of subadditivity violation by VaR (see Acerbi 2003, Examples 2.15 and 4.4) allow us to display the connection between the coherence of a risk measure and the ~ 7! rðPð~ convexity of risk surfaces. By risk surface, we mean the function w w ÞÞ which maps P ~ of the portfolio Pð~ w X onto the risk rðPð~ w ÞÞ of the the vector of weights w wÞ ¼ i i i portfolio. The problem of r-portfolio optimization amounts to a global search for minima of the surface. An elementary consequence of coherency is the convexity of risk surfaces (see Figure 12.1) r coherent ) rðPð~ w ÞÞ convex:

ð12:5Þ

This immediate result tells us that risk-optimization—if we just define carefully our variables—is an intrinsically convex problem. This bears enormous practical consequences, because the border between convex and non-convex optimization in fact delimits solvable and unsolvable problems when things are complex enough, whatever supercomputer you may have. In examples, VaR exhibits non-convex risk surfaces, infested with local minima, which can easily be recognized to be just artefacts of the chosen (noncoherent) risk measure (see Figure 12.2). In the same examples, thanks to convexity, any CRM displays conversely a single global minimum, which corresponds undoubtedly to the correct optimal portfolio, as can be shown by symmetry arguments. The lesson we learn is that adopting a non-coherent measure as a decision-making tool for asset allocation means choosing to face formidable (and often unsolvable) computational problems related to the minimization of risk surfaces plagued by a plethora of risk
nonsensical local minima. As a matter of fact we are persuaded that no bank in the world has actually ever performed a true VaR minimization in its portfolios, if we exclude multivariate Gaussian frameworks a´ la Riskmetrics where VaR is actually just a disguised version of standard deviation and hence convex.

/

262 j CHAPTER 12 0.1 ES TCE VaR

0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0

0

50

100

150

200

250

300

350

400

FIGURE 12.2 The risk of an equally weighted portfolio of i.i.d. defaultable bonds, as a function of the number of bonds N. The global minimum is expected at N   by diversification principle and symmetry arguments. Notice however that VaR and TCE (Tail Conditional Expectation (Artzner et al. 1997)), being non-coherent, display also an infinite number of nonsensical local minima. Do not expect convexity for the ES plot because it is not a ‘risk surface’ (which at N  400 would be a 400-dimensional hypersurface in R401). From Acerbi (2003) Example 4.4.

Nowadays, sacrificing the huge computational advantage of convex optimization only for the sake of VaR fanaticism is pure masochism.

12.3

THE OBJECTIVIST PARADIGM

The general representation of CRMs is well known (Artzner et al. 1999; Delbaen 2000). Any CRM rF is in one-to-one correspondence with a family F of probability measures P. The formula is strikingly simple rF ðXÞ ¼ sup E P ½X :

ð12:6Þ

P2F

But this representation is of little help for a risk manager. It gives him too much freedom. And more importantly, it generates a sort of philosophical impasse as it assumes an intrinsically subjectivist point of view that is opposite to the typical risk manager’s philosophy, which is objectivist. The formula defines the CRM rF as the worst case expected loss of the portfolio in a family F of ‘parallel universes’ P. Objectivists are, in a word, those statisticians who believe that a unique real probability measure of future events must necessarily exist somewhere, and whose principal aim is to try to estimate it empirically. Subjectivists on the contrary, are intransigent statisticians who, starting from the observation that even if this real probability measure existed it would be unknowable, simply prefer to give up considering this concept at all and think of probability measures as mere mathematical instruments. Representation (12.6) is manifestly subjectivist as it is based on families of probability measures.

COHERENT MEASURES OF RISK IN EVERYDAY MARKET PRACTICE

j

263

Risk Managers are objectivists because the algorithm they use to assess capital adequacy via VaR is intrinsically objectivist. We can in fact split this process into two clearly distinct steps: 1. model the probability distribution of your portfolio; 2. compute VaR on this distribution. An overwhelmingly larger part of the computational effort (data mining, multivariate risk-factors distribution modelling, asset pricing, . . .) is done in step 1, which has no relation with VaR and is just an objectivist paradigm. The computation of VaR, given the distribution, is typically a single last code line. Hence, in this scheme, replacing VaR with any other CRM is immediate, but it is clear that for this purpose it is necessary to identify those CRMs that fit the objectivist paradigm. If we look for something better than VaR we cannot forget that despite its shortcomings, this risk measure brought a real revolution into risk management practice thanks to some features, which were innovative at the time of its advent, and that nobody would be willing to give up today.





universality (VaR applies to risks of any nature); globality (VaR condenses multiple risks into a single figure); probability (VaR contains probabilistic information on the measured risks); right units of measure (VaR is simply expressed in ‘lost money’).

The last two features explain why VaR is worshipped by any firm’s boss, whose daily refrain is: ‘how much money do we risk and with what probability?’ Remember that risk sensitivities (aka ‘greeks,’ namely partial derivatives of the portfolio value with respect to a specific risk factor) do not share any of the above features and you will immediately understand why VaR became so popular. As a matter of fact a bank’s greeks-based risk report is immensely more cumbersome and less communicative than a VaR-based one. But if we look more closely at the features that made the success of VaR, we notice that in fact they have nothing to do with VaR in particular, but rather with the objectivist paradigm above. In other words, if in step 2 above, we replace VaR with any sensible risk measure defined in terms of some monetary statistics of the portfolio distribution, we automatically preserve these features. That is why looking for CRMs that fit the objectivist paradigm is so crucial. In our opinion, the real lasting heritage of VaR in the development of the theory and practice of risk management is precisely the very fact that it served to introduce for the first time the objectivist paradigm into the market practice. Risk managers started to plot the distribution of their portfolio’s values and learned to fear its left tail thanks to the lesson of VaR.

12.4

ESTIMABILITY

The property that characterizes the subset of those CRMs that fit the objectivist paradigm is law invariance, first studied in this context by Kusuoka (2001). A measure of risk r is said to be law-invariant (LI) if it is a functional of the portfolio’s distribution function

264 j CHAPTER 12

FX( ×) only. The concept of law invariance therefore can be defined only with reference to a single chosen probability space, r law invariant , rðXÞ ¼ r½FX ðÞ

ð12:7Þ

r law invariant , ½FX ðÞ ¼ FY ðÞ ) rðXÞ ¼ rðY Þ:

ð12:8Þ

or equivalently

It is easy to see that law invariance means estimability from empirical data. Theorem 12.1: r law invariant , r estimable

ð12:9Þ

Proof ( 0: It is easy to see that f+ is positive homogeneous convex, non-decreasing, lsc, and satisfies fðZÞ > Z for all Z 6¼ 0, but is not finite. Example 13.1.4 (Higher moment coherent risk measures): Let X ¼ Lp ðO; F;  PÞ,p and 1=p 1 þ . for some 0 B a B 1 consider fðXÞ ¼ ð1  aÞ kðXÞ kp ; where kXkp ¼ EjXj Clearly, f satisfies the conditions of Theorem (13.1), thus we can define a family of higher-moment coherent risk measures (HMCR) as   1 þ HMCRp;a ðXÞ ¼ min Z þ ð1  aÞ ðX  ZÞ p ; Z2R

p  1; a 2 ð0; 1Þ:

ð13:4Þ

From the fact that kXkp  kXkq for 1  p < q, it immediately follows that the HMCR measures are monotonic with respect to the order p: HMCR p;a ðX Þ  HMCR q;a ðXÞ

for

p 1:

ð13:12Þ

A formal derivation of equality (13.12) can be carried out using the techniques employed in Rockafellar and Uryasev (2002) to establish formula (13.11). Although the optimal hp,a(X) in (12) is determined implicitly, Theorem 13.2 ensures that it has properties similar to those of VaRa (monotonicity, positive homogeneity, etc). Moreover, by plugging relation (13.12) with p
2 into (13.6), the second-moment coherent risk measure (SMCR) can be presented in the form that involves only the first moment of losses in the tail of the distribution:   SMCRa ðXÞ ¼ Z2;a ðXÞ þ ð1  aÞ2 ðX  Z2;a ðXÞÞþ 1 þ 2  ¼ Z2;a ðXÞ þ ð1  aÞ E X  Z2;a ðXÞ :

ð13:13Þ

HIGHER MOMENT COHERENT RISK MEASURES

j

279

Note that in (13.13) the second-moment information is concealed in the h2, a(X). Further, by taking a CVaR measure with the confidence level a+
2aa2, one can write  þ 1 E X  Zsmcr 1  a  þ 1 E X  Zcvar  Zcvar þ 1  a ¼ CVaRa ðXÞ;

SMCRa ðXÞ ¼ Zsmcr þ

ð13:14Þ

where hsmcr
hp,a(X) as in (13.12) with p
2, and Zcvar ¼ VaRa ðXÞ; note that the inequality in (13.14) holds due to the fact that Zcvar ¼ VaRa ðXÞ minimizes the expression Z þ ð1  a Þ1 EðX  ZÞþ . In other words, with the above selection of a and a+, the expressions for SMCRa ðXÞ and CVaRa ðXÞ differ only in the choice of h that delivers the minimum to the corresponding expressions (13.6) and (13.2). For Conditional Value-atRisk, it is the a-quantile of the distribution of X, whereas the optimal h2,a(X) for the SMCR measure incorporates the information on the second moment of losses X. Example 13.2.3 (HMCR as tail risk measures): It is easy to see that the HMCR family are tail risk measures, namely that 0 B a 1 B a 2 B 1 implies Zp;a1 ðXÞ  Zp;a2 ðXÞ, 1 þ in addition, one has where Zp;ai ¼ Arg minZ fZ þ ð1  ai Þ kðX  ZÞ kp g; lima!1 Zp;a ðXÞ ¼ ess:sup X, at least when ess.sup X is finite (see the appendix). These properties puts the HMCR family in a favorable position comparing to the coherent measures of type (13.8) (Fischer 2003; Rockafellar et al. 2006), where the ‘tail cutoff ’ point, about which the partial moments are computed, is always fixed at EX. In contrast to (13.8), the location of tail cutoff in the HMCR measures is determined by the optimal hp, a(X) and is adjustable by means of the parameter a. In a special case, for example, the HMCR measures (13.4) can be reduced to form (13.8) with b ¼ ð1  ap Þ1 > 1 by selecting ap according to (13.12) as  p1 þ þ ; p > 1: ap ¼ 1  kðX  EXÞ kp1 kðX  EXÞ kp 13.2.1.1

Representations based on relaxation of (A3) Observe that formula (1) in Theorem 13.1

is analogous to the operation of infimal convolution, well known in convex analysis: ð f & gÞðxÞ ¼ inf f ðx  yÞ þ g ðyÞ: y

Continuing this analogy, consider the operation of right scalar multiplication ðfZÞðXÞ ¼ ZfðZ1 XÞ;

Z  0;

where for h
0 we set ðf0ÞðXÞ ¼ ðf0þ ÞðX Þ: If f is proper and convex, then it is known that (fh)(X) is a convex proper function in h ] 0 for any X  dom f (see, e.g. Rockafellar 1970). Interestingly enough, this fact can be pressed into service to formally define a coherent risk measure as

280 j CHAPTER 13

  rðXÞ ¼ inf Zf Z1 X ;

ð13:15Þ

Z0

if function f, along with some technical conditions similar to those of Theorem 13.1, satisfies axioms (A1), (A2?) and (A4). Note that excluding the positive homogeneity (A3) from the list of properties of f denies also its convexity, thus one must replace (A2) with (A2?) to ensure convexity in (13.15). In the terminology of convex analysis the function rðXÞ defined by (13.15) is known as the positively homogeneous convex function generated by f. Likewise, by direct verification of conditions (A1) (4) it can be demonstrated that /

rðX Þ ¼ sup ZfðZ1 X Þ;

ð13:16Þ

Z>0

is a proper coherent risk measure, provided that f(X) satisfies (A1), (A2) and (A4). By (C1), axioms (A1) and (A2) imply that f(0)
0, which allows one to rewrite (13.16) as rðXÞ ¼ sup Z>0

fðZX þ 0Þ  fð0Þ ¼ f0þ ðXÞ; Z

ð13:17Þ

where the last equality in (13.17) comes from the definition of the recession function (Rockafellar 1970; Za˘linescu 2002). 13.2.2

Implementation in Stochastic Programming Problems

The developed results can be efficiently applied in the context of stochastic optimization, where the random outcome X ¼ Xðx; oÞ can be considered as a function of the decision vector x  Rm, convex in x over some closed convex set C ƒ Rm. Firstly, representation (1) allows for efficient minimization of risk in stochastic programs. For a function f that complies with the requirements of Theorem 13.1, denote   Fðx; ZÞ ¼ Z þ f Xðx; oÞ  Z

and

  RðxÞ ¼ r Xðx; oÞ ¼ min Fðx; ZÞ: Z2R

ð13:18Þ

Then, clearly,   min r Xðx; oÞ x2C

()

min

ðx;ZÞ 2 CR

Fðx; ZÞ;

ð13:19Þ

in the sense that both problems have the same optimal objective values and optimal vector x+. The last observation enables seamless incorporation of risk measures into stochastic programming problems, thereby facilitating the modelling of risk-averse preferences. For example, a generalization of the classical 2-stage stochastic linear programming (SLP) problem (see, e.g. Pre´kopa 1995; Birge and Louveaux 1997) where the outcome of the second-stage (recourse) action is evaluated by its risk rather than the expected value, can be formulated by replacing the expectation operator in the second-stage problem with a coherent risk measure R:

HIGHER MOMENT COHERENT RISK MEASURES

min x0



>

>

j

281



c x þ R min qðoÞ yðoÞ y0

s:t: Ax ¼ b;

TðoÞ x þ WðoÞ yðoÞ ¼ hðoÞ:

ð13:20Þ

Note that the expectation operator is a member of the class of coherent risk measures defined by (A1)–(A4), whereby the classical 2-stage SLP problem is a special case of (13.20). Assuming that the risk measure R above is amenable to representation (1) via some function f, problem (13.20) can be implemented by virtue of Theorem 13.1 as   > min c> x þ Z þ f qðoÞ yðoÞ  Z s:t: Ax ¼ b; TðoÞ x þ WðoÞ yðoÞ ¼ hðoÞ; x  0;

yðoÞ  0;

Z 2 R;

with all the distinctive properties of the standard SLP problems (e.g. the convexity of the recourse function, etc.) being preserved. Secondly, representation (1) also admits implementation of risk constraints in stochastic programs. Namely, let g(x) be a function that is convex then, the following two problems are equivalent, as demonstrated by Theorem 13.3 below: min f gðxÞ j RðxÞ  cg; x2C

min

ðx;ZÞ 2 CR

f gðxÞ j Fðx; ZÞ  cg:

ð13:21aÞ

ð13:21bÞ

Theorem 13.3 Optimization problems (13.21a) and (13.21b) are equivalent in the sense that they achieve minima at the same values of the decision variable x and their optimal objective values coincide. Further, if the risk constraint in (13.21a) is binding at optimality, ðx; ZÞ achieves the minimum of (13.21b) if and only if x+ is an optimal solution of (13.21a) and Z 2 arg minZ Fðx; ZÞ.   In other words, one can implement the risk constraint r Xðx; oÞ  c by using representation (1) for the risk measure r with the infimum operator omitted. 13.2.2.1

HMCR measures in stochastic programming problems The introduced higher moment

coherent risk measures can be incorporated in stochastic programming problems via conic constraints of order p  1. Namely, let {v1, . . . , vJ} ⁄ V where Pfoj g ¼ pj 2 ð0; 1Þ be the scenario set of a stochastic programming model. Observe that objective or  a HMCR-based  constraint can be implemented via the constraint HMCRp;a Xðx; oÞ , with u being either a variable or a constant, correspondingly. By virtue of Theorem 13.3, the latter constraint admits a representation by the set of inequalities u  Z þ ð1  aÞ1 t;

ð13:22aÞ

282 j CHAPTER 13

1=p  t  w1p þ    þ wJp ; 1=p 

wj  pj

 Xðx; oj Þ  Z ;

wj  0;

ð13:22bÞ

j ¼ 1; . . . ; J ;

j ¼ 1; . . . ; J :

ð13:22cÞ

ð13:22dÞ

Note that the convexity of X as a function of the decision variables x implies convexity of (13.22c), and, consequently, convexity of the set (13.22). Constraint (13.22b) defines a J  1-dimensional cone of order p, and is central to practical implementation of constraints (13.22) in mathematical programming models. In the special case of p
2, it represents a second-order (quadratic) cone in RJ1, and well-developed methods of second-order cone programming (SOCP) can be invoked to handle the constructions of type (13.22). In the general case of p  (1, ), the p-order cone within the positive orthant 

w1p þ    þ wJp

1=p

 t;

t; wj  0; j ¼ 1; . . . ; J ;

ð13:23Þ

can be approximated by linear inequalities when J
2d. Following Ben-Tal and Nemirovski (1999), the 2d  1-dimensional p-order conic constraint (13.23) can be represented by a set of 3-dimensional p-order conic inequalities h

ðk1Þ p

w2j1

ðdÞ

 ðk1Þ p i1=p ðkÞ þ w2j  wj ;

j ¼ 1; . . . ; 2dk ;

k ¼ 1; . . . ; d;

ð13:24Þ

ð0Þ

where w1  t and wj  w j ð j ¼ 1; . . . ; 2d Þ. Each of the 3-dimensional p-order cones (13.24) can then be approximated by a set of linear equalities. For any partition 0  a0 < a1 <    < am  p=2 of the segment [0, p/2], an internal approximation of the p-order cone in the positive orthant of R3 1=p  x3  xp1 þ xp2 ;

x1 ; x2 ; x3  0;

ð13:25Þ

can be written in the form   x3 sin2=p aiþ1 cos2=p ai  cos2=p aiþ1 sin2=p ai      x1 sin2=p aiþ1  sin2=p ai þ x2 cos2=p ai  cos2=p aiþ1 ; i ¼ 0; . . . ; m  1;

ð13:26aÞ

and an external approximation can be constructed as  ðp1Þ=p x3 cosp ai þ sinp ai  x1 cosp1 ai þ x2 sinp1 ai ;

i ¼ 0; . . . ; m:

ð13:26bÞ

HIGHER MOMENT COHERENT RISK MEASURES

j

283

For example, the uniform partition ai
p i/2m (i
0 , . . . , m) generates the following approximations of a 3-dimensional second-order cone: x3 cos

p pð2i þ 1Þ pð2i þ 1Þ  x1 cos þ x2 sin ; 4m 4m 4m x3  x1 cos

13.2.3

pi pi þ x2 sin ; 2m 2m

i ¼ 0; . . . ; m  1;

i ¼ 0; . . . ; m:

Application to Deviation Measures

Since being introduced in Artzner et al. (1999), the axiomatic approach to construction of risk measures has been repeatedly employed by many authors for the development of other types of risk measures tailored to specific preferences and applications (see Acerbi 2002; Rockafellar et al. 2006; Ruszczyn´ski and Shapiro 2006). In this subsection we consider deviation measures as introduced by Rockafellar et al. (2006). Namely, deviation measure is a mapping D : X  [0, ] that satisfies (D1) (D2) (D3) (D4)

D(X)  0 for any non-constant X  X , whereas D(X)
0 for constant X, D(X  Y) 5 D(X)  D(Y) for all X, Y  X, 0 5 l 5 1, D(lX)
lD(X) for all X  X , l  0, D (X  a)  D(X) for all X  X , a  R.

Again, from axioms (D1) and (D2) follows convexity of D(X). In Rockafellar et al. (2006) it was shown that deviation measures that further satisfy (D5) D(X) 5 ess.sup X  EX for all X  X , are characterized by the one-to-one correspondence DðXÞ ¼ RðX  EXÞ

ð13:27Þ

with expectation-bounded coherent risk measures, i.e. risk measures that satisfy (A1) (A4) and an additional requirement (A5) R(X)  EX, for all non-constant X  X , whereas RðXÞ ¼ EX for all constant X. Using this result, it is easy to provide an analogue of formula (1) for deviation measures. /

Theorem 13.4 Let function f : X  R satisfy axioms (A1) (A3), and be a lsc function such that f (X)  EX for all X " 0. Then the optimal value of the stochastic programming problem /

DðX Þ ¼ EX þ inf fZ þ fðX  ZÞg Z

ð13:28Þ

is a deviation measure, and the infimum is attained for all X, so that infh in (13.28) may be replaced by minh R. Given the close relationship between deviation measures and coherent risk measures, it is straightforward to apply the above results to deviation measures.

284 j CHAPTER 13

13.2.4

Connection with Utility Theory and Second-Order Stochastic Dominance

As has been mentioned in Section 13.1, considerable attention has been devoted in the literature to the development of risk models and measures compatible with the utility theory of von Neumann and Morgenstern (1944), which represents one of the cornerstones of the decision-making science. The vNM theory argues that when the preference relation ‘’ of the decision-maker satisfies certain axioms (completeness, transitivity, continuity and independence), there exists a function u : R  R, such that an outcome X is preferred to outcome Y (‘X Y’) if and only if E½uðXÞ  E½uðY Þ. If the function u is non-decreasing and concave, the corresponding preference is said to be risk averse. Rothschild and Stiglitz (1970) have bridged the vNM utility theory with the concept of the second-order stochastic dominance by showing that X dominating Y by the second-order stochastic dominance, X SSD Y , is equivalent to the relation E½uðXÞ  E½uðY Þ holding true for all concave non-decreasing functions u, where the inequality is strict for at least one such u. Recall that a random outcome X dominates outcome Y by the second-order stochastic dominance if Z

z

P½X  t dt  1

Z

z

P½Y  tdt for all z 2 R: 1

Since coherent risk measures are generally inconsistent with the second-order stochastic dominance (see an example in De Giorgi (2005)), it is of interest to introduce risk measures that comply with this property. To this end, we replace the monotonicity axiom (A1) in the definition of coherent risk measures by the requirement of SSD isotonicity (Pflug 2000; De Giorgi 2005): ðXÞ SSD ðY Þ ) RðXÞ  RðY Þ: Namely, we consider risk measures R: X 7! R that satisfy the following set of axioms:3 (A1?) SSD isotonicity  : ðXÞ SSD ðY  Þ ) RðXÞ  RðY Þ for X; Y 2 X ; (A2?) convexity : R lX þ ð1  lÞY  lRðXÞ þ ð1  lÞRðY Þ; X; Y 2 X ; 0  l  1; (A3) positive homogeneity: RðlXÞ ¼ lRðXÞ; X 2 X ; l > 0; (A4) translation invariance : RðX þ aÞ ¼ RðXÞ þ a; X 2 X ; a 2 R: Note that unlike the system of axioms (A1) (A4), the above axioms, and in particular (A1?), require X and Y to be integrable, i.e. one can take the space in (A1?) (A4) to be L1 (for a discussion of topological properties of sets defined by stochastic dominance relations, see, e.g. Dentcheva and Ruszczyn´ski (2004). /

/

3

See Mansini et al. (2003) and Ogryczak and Opolska-Rutkowska (2006) for conditions under which SSD-isotonic measures also satisfy the coherence properties.

HIGHER MOMENT COHERENT RISK MEASURES

j

285

Again, it is possible to develop an analogue of formula (1), which would allow for construction of risk measures with the above properties using functions that comply with (A1?), (A2?) and (A3). Theorem 13.5 Let function f: X  R satisfy axioms (A1?), (A2?), and (A3), and be a lsc function such that f(h) h for all real h " 0. Then the optimal value of the stochastic programming problem rðXÞ ¼ min Z þ fðX  ZÞ Z2R

ð13:29Þ

exists and is a proper function that satisfies (A1?), (A2?), (A3) and (A4). Obviously, by solving the risk-minimization problem   min r Xðx; oÞ ; x2C

where r is a risk measure that is both coherent and SSD-compatible in the sense of (A1?), one obtains a solution that is SSD-efficient, i.e. acceptable to any risk-averse rational utility maximizer, and also bears the lowest risk in terms of coherence preference metrics. Below we illustrate that functions f satisfying the conditions of Theorem 13.5 can be easily constructed in the scope of the presented approach. Example 13.5.1 Let fðXÞ ¼ E½uðXÞ, where u : R  R is a convex, positively homogeneous, non-decreasing function such that u(h) h for all h " 0. Obviously, function f(X) defined in this way satisfies the conditions of Theorem 13.5. Since u( h) is concave and non-decreasing, one has that E½uðXÞ  E½uðY Þ, and, consequently, f(X) 5 f(Y), whenever ðXÞ SSD ðY Þ. It is easy to see that, for example, function f of the form þ



fðXÞ ¼ aEðXÞ  bEðXÞ ;

a 2 ð1; þ1Þ; b 2 ½0; 1Þ;

satisfies the conditions of Theorem 13.5. Thus, in accordance with Theorems 13.1 and 13.5, the coherent risk measure Ra,b (3) is also consistent with the second-order stochastic dominance. A special case of (3) is Conditional Value-at-Risk, which is known to be compatible with the second-order stochastic dominance (Pflug 2000). Example 13.5.2 (Higher moment coherent risk measures) SMCR and, in general, the family of higher-moment coherent risk measures constitute another example of risk measures that are both coherent and compatible with the second-order stochastic þ p dominance. Indeed, function uðZÞ ¼ ððZÞ Þ is convex and non-decreasing, whence 1=p 1=p ðE½uðXÞÞ  ðE½uðY ÞÞ for any ðXÞ SSD ðY Þ. Thus, the HMCR family, defined by 1 þ (29) with fðXÞ ¼ ð1  aÞ kðXÞ kp ,

286 j CHAPTER 13

  1 þ HMCRp;a ðX Þ ¼ min Z þ ð1  aÞ ðX  ZÞ p ; Z2R

p  1;

is both coherent and SSD-compatible, by virtue of Theorems 13.1 and 13.5. Implementation of such a risk measure in stochastic programming problems enables one to introduce risk preferences that are consistent with both concepts of coherence and second-order stochastic dominance. The developed family of higher-moment risk measures (4) possesses all the outstanding properties that are sought after in the realm of risk management and decision making under uncertainty: compliance with the coherence principles, amenability for an efficient implementation in stochastic programming problems (e.g. via second-order cone programming), and compatibility with the second-order stochastic dominance and utility theory. The question that remains to be answered is whether these superior properties translate into an equally superior performance in practical risk management applications. The next section reports a pilot study intended to investigate the performance of the HMCR measures in real-life risk management applications. It shows that the family of HMCR measures is a promising tool for tailoring risk preferences to the specific needs of decision-makers, and can be compared favourably with some of the most widely used risk management frameworks.

13.3

PORTFOLIO CHOICE USING HMCR MEASURES: AN ILLUSTRATION

In this section we illustrate the practical utility of the developed HMCR measures on the example of portfolio optimization, a typical testing ground for many risk management and stochastic programming techniques. To this end, we compare portfolio optimization models that use the HMCR measures with p
2 (SMCR) and p
3 against portfolio allocation models based on two well-established, and theoretically as well as practically proven methodologies, the Conditional Value-at-Risk measure and the Markowitz MeanVariance framework. This choice of benchmark models is further supported by the fact that the HMCR family is similar in the construction and properties to CVaR (more precisely, CVaR is a HMCR measure with p
1), but, while CVaR measures the risk in terms of the first moment of losses residing in the tail of the distribution, the SMCR measure quantifies risk using the second moments, in this way relating to the MV paradigm. The HMCR measure with p
3 demonstrates the potential benefits of using higher-order tail moments of loss distributions for risk estimation. 13.3.1

Portfolio Optimization Models and Implementation

The portfolio selection models employed in this case study have the general form min x

Rðr> xÞ

ð13:30aÞ

HIGHER MOMENT COHERENT RISK MEASURES

j

287

s:t: e> x ¼ 1;

ð13:30bÞ

Er> x  r0 ;

ð13:30cÞ

x  0;

ð13:30dÞ

where x
(x1 , . . . , xn)i is the vector of portfolio weights, r
(r1, . . ., rn)i is the random vector of assets’ returns, and e
(1, . . . , 1)i. The risk measure R in (13.30a) is taken to be either SMCR (6), HMCR with p
3 (4), CVaR (2), or variance s2 of the negative portfolio return, rix
X. In the above portfolio optimization problem, (13.30b) represents the budget constraint, which, together with the no-short-selling constraint (13.30d) ensures that all the available funds are invested, and (13.30c) imposes the minimal required level r0 for the expected return of the portfolio. We have deliberately chosen not to include any additional trading or institutional constraints (transaction costs, liquidity constraints, etc.) in the portfolio allocation problem (13.30) so as to make the effect of risk measure selection in (13.30a) onto the resulting portfolio rebalancing strategy more marked and visible. Traditionally to stochastic programming, the distribution of random return ri of asset i is modelled using a set of J discrete equiprobable scenarios {ri1, . . . , r i J}. Then, optimization problem (13.30) reduces to a linear programming problem if CVaR is selected as the risk measure R in (13.30a) (see, for instance, Rockafellar and Uryasev 2000; Krokhmal et al. 2002a). Within the Mean-Variance framework, (13.30) becomes a convex quadratic optimization problem with the objective >

Rðr xÞ ¼

n X

sik xi xk ;

where

i;k¼1 J J 1 X 1X sik ¼ ðrij  ri Þðrkj  rk Þ; ri ¼ rij : J  1 j¼1 J j¼1

ð13:31Þ

In the case of RðXÞ ¼ HMCRp;a ðXÞ; problem (13.30) transforms into a linear programming problem with a single p-order cone constraint (13.32e): Zþ

min s:t:

1

1

1  a J 1=p

n X

t

ð13:32aÞ

x i ¼ 1;

ð13:32bÞ

rij xi  r0 ;

ð13:32cÞ

i¼1 J n X 1X

J wj  

j¼1

n X i¼1

i¼1

rij xi  Z;

j ¼ 1; . . . ; J ;

ð13:32dÞ

288 j CHAPTER 13

1=p  t  w1p þ    þ wJp ;

ð13:32eÞ

xi  0;

i ¼ 1; . . . ; n;

ð13:32f Þ

wj  0;

j ¼ 1; . . . ; J :

ð13:32g Þ

When p
2, i.e. R is equal to SMCR, (32) reduces to a SOCP problem. In the case when R is selected as HMCR with p
3, the 3rd-order cone constraint (13.32e) has been approximated via linear inequalities (13.26a) with m
500, thereby transforming problem (13.32) to a LP. The resulting mathematical programming problems have been implemented in C, and we used CPLEX 10.0 for solving the LP and QP problems, and MOSEK 4.0 for solving the SOCP problem. 13.3.2

Set of Instruments and Scenario Data

Since the introduced family of HMCR risk measures quantifies risk in terms of higher tail moments of loss distributions, the portfolio optimization case studies were conducted using a data set that contained return distributions of fifty S&P 500 stocks with the socalled ‘heavy tails.’ Namely, for scenario generation we used 10-day historical returns over J
1024 overlapping periods, calculated using daily closing prices from 30 October 1998 to 18 January 2006. From the set of S&P 500 stocks (as of January 2006) we selected n
50 instruments by picking the ones with the highest values of kurtosis of biweekly returns, calculated over the specified period. In such a way, the investment pool had the average kurtosis of 51.93, with 429.80 and 17.07 being the maximum and minimum kurtosis, correspondingly. The particular size of scenario model, J
1024
210, has been chosen so that the linear approximation techniques (13.26) can be employed for the HMCR measure with p
3. 13.3.3

Out-of-Sample Simulations

The primary goal of our case study is to shed light on the potential ‘real-life’ performance of the HMCR measures in risk management applications, and to this end we conducted the so-called out-of-sample experiments. As the name suggests, the out-of-sample tests determine the merits of a constructed solution using out-of-sample data that have not been included in the scenario model used to generate the solution. In other words, the out-of-sample setup simulates a common situation when the true realization of uncertainties happens to be outside of the set of the ‘expected,’ or ‘predicted’ scenarios (as is the case for most portfolio optimization models). Here, we employ the out-of-sample method to compare simulated historic performances of four self-financing portfolio rebalancing strategies that are based on (13.30) with R chosen either as a member of the HMCR family with a
0.90, namely, CVaR0.90(×), SMCR0.90( ×) HMCR3,0.90( ×), or as variance s2( ×). It may be argued that in practice of portfolio management, instead of solving (13.30), it is of more interest to construct investment portfolios that maximize the expected return subject to risk constraint(s), e.g.

HIGHER MOMENT COHERENT RISK MEASURES

n o  max Er> x  Rðr> xÞ  c0 ; e> x ¼ 1 : x0

j

289

ð13:33Þ

Indeed, many investment institutions are required to keep their investment portfolios in compliance with numerous constraints, including constraints on risk. However, our main point is to gauge the effectiveness of the HMCR risk measures in portfolio optimization by comparing them against other well-established risk management methodologies, such as the CVaR and MV frameworks. And since these risk measures yield risk estimates on different scales, it is not obvious which risk tolerance levels c0 should be selected in (13.33) to make the resulting portfolios comparable. Thus, to facilitate a fair ‘apple-to-apple’ comparison, we construct self-financing portfolio rebalancing strategies by solving the risk-minimization problem (13.30), so that the resulting portfolios all have the same level r0 of expected return, and the success of a particular portfolio rebalancing strategy will depend on the actual amount of risk borne by the portfolio due to the utilization of the corresponding risk measure. The out-of-sample experiments have been set up as follows. The initial portfolios were constructed on 11 December 2002 by solving the corresponding variant of problem (13.30), where the scenario set consisted of 1024 overlapping bi-weekly returns covering the period from 30 October 1998 to 11 December 2002. The duration of the rebalancing period for all strategies was set at two weeks (ten business days). On the next rebalancing date of 26 December 2002,4 the 10-day out-of-sample portfolio returns, ^r> x, were observed for each of the three portfolios, where ^r is the vector of out-of-sample (11 December 2002–26 December 2002) returns and x+ is the corresponding optimal portfolio configuration obtained on 11 December 2002. Then, all portfolios were rebalanced by solving (13.30) with an updated scenario set. Namely, we included in the scenario set the ten vectors of overlapping biweekly returns that realized during the ten business days from 11 December 2002 to 26 December 2002, and discarded the oldest ten vectors from October–November of 1998. The process was repeated on 26 December 2002, and so on. In such a way, the out-of-sample experiment consisted of 78 biweekly rebalancing periods covering more than three years. We ran the out-of-sample tests for different values of the minimal required expected return r0 and typical results are presented in Figures 13.1 to 13.3. Figure 13.1 reports the portfolio values (in percent of the initial investment) for the four portfolio rebalancing strategies based on (13.30) with SMCR0.90(×) CVaR0.90( ×), variance s2( ×), and HMCR3,0.90( ×) as R( ×), and the required level r0 of the expected return being set at 0.5). One can observe that in this case the clear winner is the portfolio based on the HMCR measure with p
3, the SMCR-based portfolio is runner-up, with the CVaR- and MV-based portfolios falling behind these two. This situation is typical for smaller values of r0; as r0 increases and the rebalancing strategies become more aggressive, the CVaR- and MV-portfolios become more competitive, while the HMCR (p
3) portfolio remains dominant most of the time (Figure 13.2). 4

Holidays were omitted from the data.

290 j CHAPTER 13 250 230

MV CVaR

210

SMCR

Portfolio Value (%)

HMCR 190 170 150 130 110

D

ec Ja 11-2 n 0 M -27- 02 ar 20 Ap 11- 03 r-2 200 Ju 3-2 3 n- 0 0 0 Ju 5-2 3 l- 0 Au 18- 03 g- 20 O 29- 03 ct 20 N 13- 03 ov 2 -2 00 Ja 4-2 3 n- 0 Fe 08-2 03 b- 0 Ap 23- 04 r- 20 M 05- 04 ay 20 0 Ju 18-2 4 l- 0 Au 01- 04 g- 200 Se 13- 4 p 20 N -27 04 ov -2 - 0 D 08- 04 ec 20 -2 0 Fe 1- 4 b- 20 M 03- 04 ar 20 M -18- 05 ay 20 Ju 02- 05 n- 20 1 0 Ju 4-2 5 l-2 00 Se 7-2 5 p- 0 O 08- 05 ct 20 D 20- 05 ec 20 - 0 05 Ja 2-2 n- 0 18 05 -2 00 6

90

FIGURE 13.1 Out-of-sample performance of conservative (r0
0.5)) self-financing portfolio rebalancing strategies based on the MV model, and CVaR, SMCR (p
2) and HMCR (p
3) measures of risk with a
0.90. 250 MV 230

CVaR SMCR

Portfolio Value (%)

210

HMCR

190 170 150 130 110

D

ec Ja 11-2 n 0 M -27- 02 ar 20 Ap 11- 03 r-2 200 Ju 3-2 3 n- 0 0 0 Ju 5-2 3 l- 0 Au 18- 03 g- 20 O 29- 03 c t 20 N 13- 03 ov 2 -2 00 Ja 4-2 3 n- 0 Fe 08-2 03 b- 0 Ap 23- 04 r- 20 M 05- 04 ay 20 0 Ju 18-2 4 l- 0 Au 01- 04 g- 200 Se 13- 4 p 20 N -27 04 ov -2 - 0 D 08- 04 ec 20 -2 0 Fe 1- 4 b- 20 M 03- 04 ar 20 M -18- 05 ay 20 Ju 02- 05 n- 20 1 0 Ju 4-2 5 l-2 00 Se 7-2 5 p- 0 O 08- 05 ct 20 D 20- 05 ec 20 -0 05 Ja 2-2 n- 0 18 05 -2 00 6

90

FIGURE 13.2 Out-of-sample performance of self-financing portfolio rebalancing strategies that have expected return r0
1.0) and are based on the MV model, and CVaR, SMCR (p
2) and HMCR (p
3) risk measures with a
0.90.

An illustration of typical behaviour of more aggressive rebalancing strategies is presented in Figure 13.3, where r0 is set at 1.3) (in the dataset used in this case study, infeasibilities in (13.30) started to occur for values of r 0  0.013). As a general trend, the HMCR (p
2, 3) and CVaR portfolios exhibit similar performance (which can be

HIGHER MOMENT COHERENT RISK MEASURES

290

j

291

MV CVaR SMCR

Portfolio Value (%)

240

HMCR

190

140

D

ec Ja 11-2 n 0 M -27- 02 ar 20 Ap 11- 03 r-2 200 Ju 3-2 3 n- 0 0 0 Ju 5-2 3 l- 0 Au 18- 03 g- 20 O 29- 03 ct 20 N 13- 03 ov 2 -2 00 Ja 4-2 3 n- 0 Fe 08-2 03 b- 0 Ap 23- 04 r- 20 M 05- 04 ay 20 0 Ju 18-2 4 l- 0 Au 01- 04 g- 200 Se 13- 4 p 20 N -27 04 ov -2 - 0 D 08- 04 ec 20 -2 0 Fe 1- 4 b- 20 M 03- 04 ar 20 M -18- 05 ay 20 Ju 02- 05 n- 20 1 0 Ju 4-2 5 l-2 00 Se 7-2 5 p- 0 O 08- 05 ct 20 D 20- 05 ec 20 -0 05 Ja 2-2 n- 0 18 05 -2 00 6

90

FIGURE 13.3 Out-of-sample performance of aggressive (r0
1.3)) self-financing portfolio rebalancing strategies based on SMCR (p
2), CVaR, MV and HMCR (p
3) measures of risk.

explained by the fact that at high values of r0 the set of instruments capable of providing such a level of expected return is rather limited). Still, it may be argued that the HMCR (p
3) strategy can be preferred to the other three, based on its overall stable performance throughout the duration of the run. Finally, to illuminate the effects of taking into account higher-moment information in estimation of risks using the HMCR family of risk measures, we compared the simulated historical performances of the SMCR measure with parameter a
0.90 and CVaR measure with confidence level a+
0.99, so that the relation a+
2a a2 holds (see the discussion in Example 13.2.2). Recall that in such a case the expressions for SMCR and CVaR differ only in the location of the optimal h+ (see (13.14)). For CVaR, the optimal h equals to VaR, and for SMCR the corresponding value of h is determined by (13.12) and depends on the second tail moment of the distribution. Figure 13.4 presents a typical outcome for mid-range values of the expected return level r0: most of the time, the SMCR portfolio dominates the corresponding CVaR portfolio. However, for smaller values of expected return (e.g. r0 5 0.005), as well as for values approaching r0
0.013, the SMCR- and CVaR-based rebalancing strategies demonstrated very close performance. This can be explained by the fact that lower values of r0 lead to rather conservative portfolios, while the values of r0 close to the infeasibility barrier of 0.013 lead to very aggressive and poorly diversified portfolios that are comprised of a limited set of assets capable of achieving this high level of the expected return. Although the obtained results are data-specific, the presented preliminary case studies indicate that the developed HMCR measures demonstrate very promising performance, and can be successfully employed in the practice of risk management and portfolio optimization. /

292 j CHAPTER 13 230

210

SMCR CVaR

Portfolio Value (%)

190

170

150

130

110

D

ec Ja 11-2 n 0 M -27- 02 ar 20 Ap 11- 03 r-2 200 Ju 3-2 3 n- 0 0 0 Ju 5-2 3 l- 0 Au 18- 03 g- 20 O 29- 03 ct 20 N 13- 03 ov 2 -2 00 Ja 4-2 3 n- 0 Fe 08-2 03 b- 0 Ap 23- 04 r- 20 M 05- 04 ay 20 0 Ju 18-2 4 l- 0 Au 01- 04 g- 200 Se 13- 4 p 20 N -27 04 ov -2 - 0 D 08- 04 ec 20 -2 0 Fe 1- 4 b- 20 M 03- 04 ar 20 M -18- 05 ay 20 Ju 02- 05 n- 20 1 0 Ju 4-2 5 l-2 00 Se 7-2 5 p- 0 O 08- 05 ct 20 D 20- 05 ec 20 -0 05 Ja 2-2 n- 0 18 05 -2 00 6

90

FIGURE 13.4 Out-of-sample performance of self-financing portfolio rebalancing strategies based on SMCR0.90 and CVaR0.99 measures of risk (r0
1.0)).

13.4

CONCLUSIONS

In this chapter we have considered modelling of risk-averse preferences in stochastic programming problems using risk measures. We utilized the axiomatic approach to construction of coherent risk measures and deviation measures in order to develop simple representations for these risk measures via solutions of specially designed stochastic programming problems. Using the developed general representations, we introduced a new family of higher-moment coherent risk measures (HMCR). In particular, we considered the second-moment coherent risk measure (SMCR), which is implementable in stochastic optimization problems using the second-order cone programming, and the third-order HMCR (p
3). The conducted numerical studies indicate that the HMCR measures can be effectively used in the practice of portfolio optimization, and compare well with the well-established benchmark models such as Mean-Variance framework or CVaR.

ACKNOWLEDGEMENT This work was supported in part by NSF grant DMI 0457473.

REFERENCES Acerbi, C., Spectral measures of risk: A coherent representation of subjective risk aversion. J. Bank. Finan., 2002, 26(7), 1487/1503. Acerbi, C. and Tasche, D., On the coherence of expected shortfall. J. Bank. Finan., 2002, 26(7), 1487 / 1503. Alizadeh, F. and Goldfarb, D., Second-order cone programming. Math. Program., 2003, 95(1), 3/51.

HIGHER MOMENT COHERENT RISK MEASURES

j

293

Artzner, P., Delbaen, F., Eber, J.M. and Heath, D., Coherent measures of risk. Math. Finan., 1999, 9(3), 203/228. Bawa, V.S., Optimal rules for ordering uncertain prospects. Rev. Financ. Stud., 1975, 2(1), 95 /121. Ben-Tal, A. and Nemirovski, A., On polyhedral approximations of the second-order cone. Math. Operat. Res., 1999, 26(2), 193/205. Ben-Tal, A. and Teboulle, M., Expected utility, penalty functions, and duality in stochastic nonlinear programming. Manag. Sci., 1986, 32(11), 1145/1466. Birge, J.R. and Louveaux, F., Introduction to Stochastic Programming, 1997 (Springer: New York). Chekhlov, A., Uryasev, S. and Zabarankin, M., Drawdown measure in portfolio optimization. Int. J. Theoret. Appl. Finan., 2005, 8(1), 13 /58. De Giorgi, E., Reward-risk portfolio selection and stochastic dominance. J. Bank. Finan., 2005, 29(4), 895/926. Delbaen, F., Coherent risk measures on general probability spaces. In Advances in Finance Stochast.: Essays in Honour of Dieter Sondermann, edited by K. Sandmann and P.J. Scho ¨ nbucher, pp. 1 /37, 2002 (Springer: Berlin). Dembo, R.S. and Rosen, D., The practice of portfolio replication: A practical overview of forward and inverse problems. Ann. Oper. Res., 1999, 85, 267/284. Dentcheva, D. and Ruszczyn´ski, A., Optimization with stochastic dominance constraints. SIAM J. Optim., 2003, 14(2), 548/566. Dentcheva, D. and Ruszczyn´ski, A., Semi-infinite probabilistic optimization: First order stochastic dominance constraints. Optimization, 2004, 53(5–6), 583/601. Duffie, D. and Pan, J., An overview of value-at-risk. J. Deri., 1997, 4, 7 /49. Fischer, T., Risk capital allocation by coherent risk measures based on one-sided moments. Insur.: Math. Econ., 2003, 32(1), 135/146. Fishburn, P.C., Stochastic dominance and moments of distributions. Math. Oper. Res., 1964, 5, 94 / 100. Fishburn, P.C., Utility Theory for Decision-Making, 1970 (Wiley: New York). Fishburn, P.C., Mean-risk analysis with risk associated with below-target returns. The Amer. Eco. Rev., 1977, 67(2), 116/126. Fishburn, P.C., Non-Linear Preference and Utility Theory, 1988 (Johns Hopkins University Press: Baltimore). Jorion, P., Value at Risk: The New Benchmark for Controlling Market Risk, 1997 (McGraw-Hill: New York). Karni, E. and Schmeidler, D., Utility theory with uncertainty. In Handbook of Mathematical Economics, Vol. IV, edited by W. Hildenbrand and H. Sonnenschein, 1991 (North-Holland: Amsterdam). Kouvelis, P. and Yu, G., Robust Discrete Optimization and Its Applications, 1997 (Kluwer Academic Publishers: Dordrecht). Krokhmal, P., Palmquist, J. and Uryasev, S., Portfolio optimization with conditional value-at-risk objective and constraints. J. Risk, 2002a, 4(2), 43 /68. Krokhmal, P., Uryasev, S. and Zrazhevsky, G., Risk management for hedge fund portfolios: A comparative analysis of linear rebalancing strategies. J. Alter. Invest., 2002b, 5(1), 10 /29. Kroll, Y., Levy, H. and Markowitz, H.M., Mean-variance versus direct utility maximization. J. Finan., 1984, 39(1), 47/61. Levy, H., Stochastic Dominance, 1998 (Kluwer Academic Publishers: Dordrecht). Mansini, R., Ogryczak, W. and Speranza, M.G., LP solvable models for portfolio optimization: A LP solvable models for portfolio optimization: A classification and computational comparison. IMA J. Manag. Math., 2003, 14(3), 187/220. Markowitz, H.M., Portfolio selection. J. Finan., 1952, 7(1), 77/91. Markowitz, H.M., Portfolio Selection, 1st, 1959 (Wiley and Sons: New York).

294 j CHAPTER 13

Ogryczak, W. and Opolska-Rutkowska, M., SSD consistent criteria and coherent risk measures, In System Modeling and Optimization. Proceedings of the 22nd IFIP TC7 Conference, Vol. 199, in the series IFIP International Federation for Information Processing, edited by F. Ceragioli, A. Dontchev, H. Furuta, M. Marti and L. Pandolfi, 2006, pp. 227 /237 (IFIP). Ogryczak, W. and Ruszczyn´ski, A., From stochastic dominance to mean-risk models: Semideviations as risk measures. Eur. J. Operat. Res., 1999, 116, 33 /50. Ogryczak, W. and Ruszczyn´ski, A., On consistency of stochastic dominance and mean-semideviation models. Math. Program., 2001, 89, 217/232. Ogryczak, W. and Ruszczyn´ski, A., Dual stochastic dominance and related mean-risk models. SIAM J. Optim., 2002, 13(1), 60 /78. Pflug, G., Some remarks on the value-at-risk and the conditional value-at-risk. In Probabilistic Constrained Optimization: Methodology and Applications, edited by S. Uryasev, 2000 (Kluwer Academic Publishers: Dordrecht). Pre´kopa, A., Stochastic Programming, 1995 (Kluwer Academic Publishers: Dordrecht). Riskmetrics, 1994 (JP Morgan: New York). Rockafellar, R.T., Convex Analysis. In the series, Princeton Mathematics, Vol. 28, 1970 (Princeton University Press: Princeton, NJ). Rockafellar, R.T. and Uryasev, S., Optimization of conditional value-at-risk. J. Risk, 2000, 2, 21 /41. Rockafellar, R.T. and Uryasev, S., Conditional value-at-risk for general loss distributions. J. Bank. Finan., 2002, 26(7), 1443/1471. Rockafellar, R.T., Uryasev, S. and Zabarankin, M., Generalized deviations in risk analysis. Finan. Stochast., 2006, 10(1), 51 /74. Roman, D., Darby-Dowman, K. and Mitra, G., Portfolio construction based on stochastic dominance and target return distributions. Math. Program., 2006, 108(1), 541/569. Rothschild, M. and Stiglitz, J., Increasing risk I: A definition. J. Econ. Theory, 1970, 2(3), 225 / 243. Ruszczyn´ski, A. and Shapiro, A., Optimization of convex risk functions. Math. Operat. Res., 2006, 31(3), 433/452. Schied, A. and Follmer, H., Robust preferences and convex measures of risk. In Advances in Finance Stochast.: Essays in Honour of Dieter Sondermann, edited by K. Sandmann and P.J. Scho ¨ nbucher, pp. 39 /56, 2002 (Springer: Berlin). Tasche, D., Expected shortfall and beyond, Working Paper. http://arxiv.org/abs/cond-mat/0203558, 2002. Testuri, C. and Uryasev, S., On relation between expected regret and conditional value-at-risk. In Handbook of Numerical Methods in Finance, edited by Z. Rachev, 2003 (Birkhauser: Boston). van der Vlerk, M.H., Integrated chance constraints in an ALM model for pension funds, Working Paper, 2003. Available online at: http://irs.ub.rug.nl/ppn/252290909. von Neumann, J. and Morgenstern, O., Theory of Games and Economic Behavior, 1953rd, 1944 (Princeton University Press: Princeton, NJ). Young, M.R., A minimax portfolio selection rule with linear programming solution. Manag. Sci., 1998, 44(5), 673/683. Za˘linescu, C., Convex Analysis in General Vector Spaces, 2002 (World Scientific: Singapore).

APPENDIX 13.A Proof of Theorem 13.1 Convexity, lower semicontinuity, and sublinearity of f in X imply that the function fX(h)
h  f(X h) is also convex, lsc, and proper in h  R for each fixed X  X . For the infimum of fX(h) to be achievable at finite h, its recession function has to be positive: fX0(91)  0, which is equivalent to fX0(j)  0, j " 0, due to the

HIGHER MOMENT COHERENT RISK MEASURES

j

295

positive homogeneity of f. By definition of the recession function (Rockafellar 1970; Za˘linescu 2002) and positive homogeneity of f, we have that the last condition holds if f (j)  j for all j " 0: fX 0þ ðxÞ ¼ lim

t!1

Z þ tx þ fðX  Z  txÞ  Z  fðX  ZÞ ¼ x þ fðxÞ: t

Hence, r(X) defined by (1) is a proper lsc function, and minimum in (1) is attained at finite h for all X  X . Below we verify that r(X) satisfies axioms (A1)–(A4). (A1) Let X 5 0. Then f(X) 5 0 as f satisfies (A1) which implies min Z þ fðX  ZÞ  0 þ fðX  0Þ  0: Z2R

(A2) For any Z  X let ZZ 2 arg minfZ þ fðZ  ZÞg  R then Z2R

rðXÞ þ rðY Þ ¼ ZX þ fðX  ZX Þ þ ZY þ fðY  ZY Þ  ZX þ ZY þ fðX þ Y  ZX  ZY Þ  ZXþY þ fðX þ Y  ZXþY Þ ¼ rðX þ Y Þ: (A3) For any fixed l  0 we have rðlXÞ ¼ min Z2R



   Z þ fðlX  ZÞ ¼ l min Z=l þ fðX  Z=lÞ ¼ lrðXÞ Z2R

ð13:A1Þ

(A4) Similarly, for any fixed a  R, o Z þ fðX þ a  ZÞ Z2R n  o ¼ a þ min ðZ  aÞ þ f X  ðZ  aÞ ¼ a þ rðXÞ:

rðX þ aÞ ¼ min

n

Z2R

ð13:A2Þ

Thus, r(X) defined by (1) is a proper coherent risk measure.

I

Proof of Theorem 13.2 Conditions on function f ensure that the set of optimal solutions of problem (1) is closed and finite, whence follows the existence of h(X) in (13.10). Property (A3) is established by noting that for any l  0 equality (13.A1) implies ZðlXÞ ¼ Arg minfZ þ fðlX  ZÞg ¼ Arg minfZ=l þ fðlX  Z=lÞg; Z2R

Z2R

from which follows that h(lX)
lh(X). Similarly, by virtue of (13.A2), we have ZðX þ aÞ ¼ Arg minfZ þ fðX þ a  ZÞg ¼ Arg minfðZ  aÞ þ fðlX  ðZ  aÞÞg; Z2R

Z2R

296 j CHAPTER 13

which leads to the sought relation (A4): h(X  a)
h (X)  a. To validate the remaining statements of the Theorem, consider f to be such that f(X)
0 for every X 5 0. Then, (C2) immediately yields f(X) ] 0 for all X  X , which proves ZðXÞ  ZðXÞ þ fðX  ZðXÞÞ ¼ rðXÞ: By the definition of h(X), we have for all X 5 0   ZðXÞ þ f X  ZðXÞ  0 þ fðX  0Þ ¼ 0;

or

  ZðXÞ  f X  ZðXÞ :

ð13:A3Þ

Assume that h(X)  0, which implies f(h(X))
0. From (A2) it follows that f(X h(X))5 f(X)  f ( h(X))
0, leading to f(Xh(X))
0, and, consequently, to h(X) 5 0 by (13.A3). The contradiction furnishes the statement of the theorem. I Proof of Theorem 13.3 Denote the feasible sets of (13.21a) and (13.21b), respectively, as S a ¼ fx 2 C j RðxÞ  c

and

S b ¼ fðx; ZÞ 2 C  R j Fðx; ZÞ  cg:

Now observe that projection PC ðS b Þ of the feasible set of (21b) onto , PC ðS b Þ ¼ fx 2 C j ðx; ZÞ 2 S b for some Z 2 Rg ¼ fx 2 C j Fðx; ZÞ  c for some Z 2 Rg; coincides with the feasible set of (13.21a): S a ¼ PC ðS b Þ: Indeed, x?  S a means that x?  C and R(x?)
minh F(x?,h) 5 c. By virtue of Theorem 13.1 there exists h?  R such that F(x?, h?)
minhF(x?, h), whence (x?, h?)  S b, and, consequently, x?  P C ðS b). If, on the other hand, xƒ  P C ðS b), then there exists hƒ  R such that (xƒ,hƒ)  S b and therefore f(xƒ, hƒ) 5 c. By definition of R( ×), R(xƒ) 5 f(xƒ, hƒ) 5 c, thus xƒ  S a. Given (13.37), it is easy to see that (13.21a) and (13.21b) achieve minima at the same values of x  C and their optimal objective values coincide. Indeed, if x+ is an optimal solution of (13.21a) then x+  S a and g(x+) 5 g(x) holds for all x  S a. By (13.37), if x  S a then there exists some h  R such (x, h)  S b. Thus, for all (x, h)  S b one has g(x+) 5 g(x), meaning that (x+, h+) is an optimal solution of (13.21b), where h+  R is such that (x+, h+)  S b. Inversely, if (x+, h+) solves (13.21b), then (x+, h+)  S b and g(x+) 5 g(x) for all (x, h)  S b. According to (13.A4), (x, h)  S b also yields x  S a, hence for all x  S a one has g(x+) 5 g(x), i.e. x+ is an optimal solution of (13.21a). Finally, assume that the risk constraint in (13.21a) is binding at optimality. If (x+, h+) achieves the minimum of (13.21b), then F(x+, h+) 5 c and, according to the above, x+ is an optimal solution of (13.21a), whence c
R(x+) 5 F(x+, h+) 5 c. From the last relation we have F(x+, h+)
R(x+) and thus h+  argminh F(x+, h). Now consider x+ that

j

HIGHER MOMENT COHERENT RISK MEASURES

297

solves (13.21a) and h+ such that h+  argminh F(x+, h). This implies that F(x+, h+)
R(x+)
c, or (x+, h+)  S b. Taking into account that g(x+) 5 g(x) for all x  S a and consequently for all (x,h)  S b, one has that (x+, h+) is an optimal solution of (13.21b). Proof of Theorem 13.4 Since formula (13.28) differs from (13.1) by the constant summand ðEXÞ, we only have to verify that R(X)
infh{h  f(X  h)} satisfies (A5). As f(X)  EX for all X " 0, we have that f(X hX)  E(X  hX) for all non-constant X  X , where hX  argminh h  f(X  h). From the last inequality it follows that hX  f(X  hX)  EX, or R(X)  EX for all non-constant X  X . Thus, D(X)  0 for I all non-constant X. For a  R, infh{h  f(a  h)}
a, whence D(a)
0. Proof of Theorem 13.5 The proof of existence and all properties except (A1?) is identical to that of Theorem 13.1. Property (A1?) follows elementarily: if ðXÞ SSD ðY Þ, then ðX þ cÞ SSD ðY þ cÞ, and consequently, f(X  c) 5 f(Y  c) for c  R, whence rðX Þ ¼ ZX þ fðX  ZX Þ  ZY þ fðX  ZY Þ  ZY þ fðY  ZY Þ ¼ rðY Þ;

ð13:A4Þ

where, as usual, hZ  arg minh{h  f(Z  h)} ƒ R, for any Z  X .

I

Example 13.2.3 (Additional details): To demonstrate the monotonicity of hp,a(X) with respect to a  (0, 1), observe that by definition of Zp;a1 ðXÞ  þ   þ  Zp;a1 ðXÞ þ ð1  a1 Þ  X  Zp;a1 p  Zp;a2 ðXÞ þ ð1  a1 Þ1  X  Zp;a2 p : ð13:A5Þ 1 

Now, assume that hp,a1(X)  hp,a2(X) for some a1 B a2, then (13.A5) yields   þ   þ       0 < Zp;a1 ðXÞ  Zp;a2 ðXÞ  ð1  a1 Þ   X  Zp;a2   X  Zp;a1  p p   þ   þ       < ð1  a2 Þ1  X  Zp;a2   X  Zp;a1  : 1

p

p

From the last inequality it follows directly that   þ  þ  1  1  Zp;a1 ðXÞ þ ð1  a2 Þ  X  Zp;a1 p < Zp;a2 ðXÞ þ ð1  a2 Þ  X  Zp;a2 p ; which contradicts the definition of Zp;a2 ðXÞ. The limiting behaviour of hp, a(X) can be verified by noting first that for 1 5 p B  lim HMCRp;a ðXÞ ¼ ess:sup X: a!1

ð13:A6Þ

298 j CHAPTER 13

Indeed, using the notation of Example 13.1.3 one has n n  o  o 1 þ 1 þ lim inf Z þ ð1  aÞ ðX  ZÞ p  inf lim Z þ ð1  aÞ ðX  ZÞ p a!1

Z

Z

a!1

¼ inf Z þ f ðX  ZÞ ¼ ess:sup X: Z

On the other hand, from the inequality (see Example 13.1.4) HMCRp;a ðXÞ  HMCRq;a ðXÞ

for

1  p < q;

and the fact that lima!1 CVaRa ðXÞ ¼ ess:sup X (see, e.g. Rockafellar et al. 2006) we obtain ess:sup X ¼ lim CVaRa ðXÞ  lim HMCRp;a ðXÞ; a!1

a!1

which verifies (13.A6). The existence of lima!1 Zp;a ðXÞ 2 R follows from the monotonicity of hp,a(X) with respect to a. Theorem 13.2 maintains that Zp;a ðXÞ  HMCRp;a ðXÞ, whence lim Zp;a ðXÞ  ess:sup X: a!1

In the case of finite ess.sup X, by rewriting (13.A6) in the form


  þ lim Zp;a ðXÞ þ ð1  aÞ ðX  Zp;a ðXÞÞ  ¼ ess:sup X; a!1

1 

p

and assuming that lima 0 1hp,a(X)
ess.supX o for some o ]0, it is easy to see that the above equality holds only in the case of o
0.

14

CHAPTER

On the Feasibility of Portfolio Optimization under Expected Shortfall ´ZARD STEFANO CILIBERTI, IMRE KONDOR and MARC ME CONTENTS 14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . 14.2 The Optimization Problem . . . . . . . . . . . . 14.3 The Replica Approach . . . . . . . . . . . . . . . 14.4 The Feasible and Infeasible Regions . . . . . 14.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . Acknowledgements . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . Appendix 14.A: The Replica Symmetric Solution

14.1

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

299 300 303 307 310 311 311 312

INTRODUCTION

A

MONG THE SEVERAL EXISTING RISK MEASURES in the context of portfolio optimization, expected shortfall (ES) has certainly gained increasing popularity in recent years. In several practical applications, ES is starting to replace the classical Value-at-Risk (VaR). There are a number of reasons for this. For a given threshold probability b, the VaR is defined so that with probability b the loss will be smaller than VaR. This definition only gives the minimum loss one can reasonably expect but does not tell anything about the typical value of that loss which can be measured by the conditional value-at-risk (CVaR, which is the same as ES for the continuous distributions that we consider here).1 We will be more precise with these definitions below. The point we want to stress here is that the VaR measure, lacking the mandatory properties of subadditivity and convexity, is not

1

See Acerbi and Tasche (2002) for the subtleties related to a discrete distribution.

299

300 j CHAPTER 14

coherent (Artzner et al. 1999). This means that summing the VaRs of individual portfolios will not necessarily produce an upper bound for the VaR of the combined portfolio, thus contradicting the holy principle of diversification in finance. A nice practical example of the inconsistency of VaR in credit portfolio management is reported by Frey and McNeil (2002). On the other hand, it has been shown (Acerbi and Tasche 2002) that ES is a coherent measure with interesting properties (Pflug 2000). Moreover, the optimization of ES can be reduced to linear programming (Rockafellar and Uryasev 2000) (which allows for a fast implementation) and leads to a good estimate for the VaR as a byproduct of the minimization process. To summarize, the intuitive and simple character, together with the mathematical properties (coherence) and the fast algorithmic implementation (linear programming), are the main reasons behind the growing importance of ES as a risk measure. In this chapter we focus on the feasibility of the portfolio optimization problem under the ES measure of risk. The control parameters of this problem are (i) the imposed threshold in probability, b and (ii) the ratio N/T between the number N of financial assets making up the portfolio and the time series length T used to sample the probability distribution of returns. It is curious, albeit trivial, that the scaling in N/T had not been explicitly pointed out before (Pafka and Kondor 2002). It was reported by Kondor et al. (2007) that, for certain values of these parameters, the optimization problem does not have a finite solution because, even if convex, it is not bounded from below. Extended numerical simulations allowed these authors to determine the feasibility map of the problem. Here, in order to better understand the root of the problem and to study the transition from a feasible to an unfeasible regime (corresponding to an ill-posed minimization problem) we address the same problem from an analytical point of view. The chapter is organized as follows. In Section 14.2 we briefly recall the basic definitions of b-VaR and b-CVaR and we show how the portfolio optimization problem can be reduced to linear programming. We introduce a ‘cost function’ to be minimized under linear constraints and we discuss the rationale for a statistical mechanics approach. In Section 14.3 we solve the problem of optimizing large portfolios under ES using the replica approach. Our results and a comparison with numerics are reported in Section 14.4, and our conclusions are summarized in Section 14.5.

14.2

THE OPTIMIZATION PROBLEM

We consider a portfolio of N financial instruments w
{w1, . . ., wN} where wi is the position of asset i. The global budget constraint fixes the sum of these numbers: we impose, for example, N X

wi ¼ N :

ð14:1Þ

i¼1

We do not stipulate any constraint on short selling, so that wi can be any negative or positive number. This is, of course, unrealistic for liquidity reasons, but considering this

ON THE FEASIBILITY OF PORTFOLIO OPTIMIZATION

j

301

case allows us to demonstrate the essence of the phenomenon. If we imposed a constraint that would render the domain of wi bounded (such as a ban on short selling, for example), this would evidently prevent the weights from diverging, but a vestige of the transition would still remain in the form of large, although finite, fluctuations of the weights, and in a large number of them sticking to the ‘walls’ of the domain. We denote the returns on the assets by x ¼ fx1 ; x2 ; . . . ; xN g and we assume that there exists an underlying probability distribution function p(x) of the returns. The loss of PN portfolio w given the returns x is ‘ðw j xÞ ¼  i¼1 wi xi , and the probability of that loss being smaller than a given threshold a is P < ðw; aÞ ¼

Z

dx pðxÞyða  ‘ðw j xÞÞ;

ð14:2Þ

where u( ×) is the Heaviside step function, equal to 1 if its argument is positive and 0 otherwise. The b-VaR of this portfolio is formally defined by b-VaRðwÞ ¼ minfa : P < ðw; aÞ bg

ð14:3Þ

(see Figure 14.1), While the CVaR (or ES, in this case) associated with the same portfolio is the average loss on the tail of the distribution, R

dx pðxÞ‘ðwjxÞyð‘ðwjxÞ  b-VaRðwÞÞ R dx pðxÞyð‘ðwjxÞ  b-VaRðwÞÞ Z 1 dx pðxÞ‘ðwjxÞyð‘ðwjxÞ  b-VaRðwÞÞ: ¼ 1b

b-CVaRðwÞ ¼

ð14:4Þ

The threshold b then represents a confidence level. In practice, the typical values of b which one considers are b
0.90, 0.95, and 0.99, but we will address the problem for any 1

P